Met A Heuristics

Marc Sevaux
Métaheuristiques
Stratégies pour l’optimisation
de la production de biens et de services
Numéro d’ordre : 04/03
H ABILITATION À D IRIGER DES R ECHERCHES
Préparée au Laboratoire d’Automatique, de Mécanique d’informatique

Industrielles et Humaines du CNRS (UMR CNRS 8530)
dans l’équipe Systèmes de Production
Marc Sevaux
Métaheuristiques
Stratégies pour l’optimisation
de la production de biens et de services
Soutenue le 1er Juillet 2004 devant le jury composé de :
Président Prof. Christian Prins

Rapporteurs Prof. Michel Gourgand
Prof. Jin-Kao Hao
Prof. Eric Taillard
Examinateurs Prof. Stéphane Dauzère-Pérès
Prof. Bernard Grabot
Directeur Prof. Christian Tahon
Table des matières
Remerciements 5
I Curriculum Vitæ 7
1 Informations générales 9
1.1 Etat Civil . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Fonction actuelle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Prime d’encadrement doctorale et de recherche . . . . . . . . . . . . 9
1.4 Parcours et formation . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Enseignement 12
2.1 Enseignements à l’UVHC . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Enseignements avant l’intégration à l’UVHC . . . . . . . . . . . . . 15
2.3 Encadrements pédagogiques . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Administration de l’enseignement . . . . . . . . . . . . . . . . . . . 18
3 Supervision de travaux d’étudiants 2e et 3e cycle 20

3.1 Thèses de doctorat . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Mémoires de DEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Projets de DESS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Projets IUP GEII . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4 Administration et animation de la recherche 24

4.1 Animation de la recherche . . . . . . . . . . . . . . . . . . . . . . . . 24
4.2 Organisation de manifestations à Valenciennes . . . . . . . . . . . . 25
4.3 Organisation de manifestations en dehors de Valenciennes . . . . . 25
4.4 Organisation / président de sessions . . . . . . . . . . . . . . . . . . 26
5 Visibilité, rayonnement et autres activités 27

5.1 Collaborations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
5.2 Appartenance à des sociétés et des groupes de recherche . . . . . . 28
5.3 Fonction d’édition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.4 Evaluation de la recherche . . . . . . . . . . . . . . . . . . . . . . . . 31
6 Contrats, projets et financements 33

6.1 Contrats industriels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.2 Projets de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
6.3 Financements obtenus . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1
7 Thématiques de recherche 36
7.1 Planification de la production . . . . . . . . . . . . . . . . . . . . . . 38
7.2 Ordonnancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
7.3 Tournées de véhicules . . . . . . . . . . . . . . . . . . . . . . . . . . 42
7.4 Autres approches ou problématiques . . . . . . . . . . . . . . . . . . 43
8 Liste des publications 48
II Synthèse scientifique 55
1 Introduction générale 57
1.1 Pourquoi les métaheuristiques ? . . . . . . . . . . . . . . . . . . . . . 57
1.2 Intensification et diversification . . . . . . . . . . . . . . . . . . . . . 57
1.3 Techniques de résolution pratique . . . . . . . . . . . . . . . . . . . 58
2 Méthodes de recherche locale 59

2.1 Méthodes de descente . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.2 Recuit simulé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.3 Recherche tabou . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.4 Recherche à voisinages variables . . . . . . . . . . . . . . . . . . . . 66
2.5 GRASP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.6 Iterated local search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2.7 Guided local search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.8 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3 Métaheuristiques à base de population 71

3.1 Algorithmes génétiques . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2 Algorithmes de colonies de fourmis . . . . . . . . . . . . . . . . . . 75
3.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4 Métaheuristiques avancées 79
4.1 Algorithmes mémétiques . . . . . . . . . . . . . . . . . . . . . . . . 79
4.2 Scatter search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3 GA|PM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5 Compléments 89
5.1 Réglages automatiques des paramètres . . . . . . . . . . . . . . . . 90
5.2 Robustesse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3 Optimisation multiobjectif . . . . . . . . . . . . . . . . . . . . . . . . 92
5.4 Optimisation continue . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6 Conclusions 93
2
7 Perspectives de recherche 96
Références 99
III Sélection de publications 109
1 European Journal of Operational Research (2003) 111
2 4OR (2004 à paraı̂tre) 112
3 Journal of Heuristics (2004 soumis) 112
4 Computers and Operations Research (2004 à paraı̂tre) 113
5 Naval Research Logistics (2003) 113
3
4
Remerciements
La rédaction de ce manuscrit est l’aboutissement de plusieurs années de tra-

vail. Ce travail lui-même n’aurait pu être mené sans l’aide de plusieurs personnes
auxquelles je souhaite exprimer ma gratitude.
Avant tout, je remercie les rapporteurs de cette dissertation, Michel Gourgand,
Jin-Kao Hao et Eric Taillard, qui ont sûrement passé de longues heures à relire et
corriger ce document. Que les autres membres du jury soient aussi remerci és.
Mes remerciements vont à la direction du LAMIH pour son accueil, et aussi
à l’ensemble de l’équipe “Systèmes de Production” et sa bonne ambiance, ainsi
qu’à Christian Tahon, son responsable, pour m’avoir fait confiance et donné une
grande liberté d’action ces dernières années.
Je tiens à exprimer ma gratitude tout particulièrement à Christian Prins, qui a
été mon professeur il y a de nombreuses années et qui a éveillé en moi le goût et
la passion de la recherche. Je lui suis aussi reconnaissant pour les nombreuses dis-
cussions productives, pour cette collaboration que nous avons entamée depuis la
fin de ma thèse et enfin, pour avoir transformé au fil des années notre partenariat
en amitié sincère.
Que Stéphane Dauzère-Pérès, mon directeur de thèse soit aussi remercié cha-
leureusement ; sous sa direction de 1996 à 1998, j’ai pu comprendre et apprendre
le métier de chercheur et une certaine éthique de la recherche. Son amitié a heu-
reusement dépassé cette période.
Je remercie aussi les professeurs qui m’ont enseignés la recherche opérationnelle
avec passion, Eric Pinson, Philippe Chrétienne, Jacques Carlier et Claude Berge
ainsi que les personnes qui ont su m’aider quand j’en avais besoin, Philippe Tho-
min, Alain Gibaud, tout particlièrement Xavier Gandibleux et bien d’autres.
Je garde une place de choix pour Kenneth Sörensen, pour tout ce que nous
avons créé et partagé ensemble, pour notre amitié ; le groupe EU/ME n’étant que
la partie émergée de l’iceberg...
Enfin je tiens à terminer cette préface en remerciant de tout cœur mon épouse
Sandrine et mes enfants, Victoria, François et Jean pour leur patience pendant
mes nombreuses absences et pour leur amour sans limite.
Marc Sevaux
Valenciennes, Mai 2004
5
6
Première partie
Curriculum Vitæ
8
1 Informations générales
1.1 Etat Civil
NOM, Prénom SEVAUX, Marc
Date et lieu Né le 25 mai 1969 à Brest (29)
de naissance
Nationalité Française
Etat civil Marié, trois enfants
Service militaire effectué en 1995-1996, en tant qu’officier du corps tech-
nique et administratif du service de santé des armées.
1.2 Fonction actuelle

Poste Maı̂tre de conférences (classe normale)
en 61e section du CNU
Adresse Université de Valenciennes et du Hainaut-Cambrésis
professionnelle UMR CNRS 8530, LAMIH / SP
Le Mont Houy, Bâtiment Jonas 2
F-59313 Valenciennes Cedex 9
France
Téléphone 03 27 51 13 26
Fax 03 27 51 13 10
Email Marc.Sevaux@univ-valenciennes.fr
Url http ://www.univ-valenciennes.fr/sp/sevaux/
Faculté Institut universitaire de technologie
de rattachement Département Organisation et Génie de la Production
(OGP)
délocalisé à Cambrai
Laboratoire LAMIH/SP : Laboratoire d’Automatique, de Mécanique,
de rattachement d’Informatique industrielles et Humaines – UMR CNRS
8530 – Equipe Systèmes de Production
1.3 Prime d’encadrement doctorale et de recherche

Titulaire de la prime d’encadrement doctorale et de recherche (PEDR)
depuis 2003.
9
1 Informations générales
1.4 Parcours et formation

Depuis Sept. 99 Maı̂tre de conférences en 61e section à l’université de Valen-
ciennes et du Hainaut-Cambrésis. Titulaire d’un poste à l’Institut Univer-
sitaire de Technologie, au département Organisation et Génie de la Produc-
tion, délocalisé à Cambrai.
Déc. 98 – Août 99 Ingénieur de recherche (contrat à durée déterminée de 9 mois)
au département Automatique et Productique de l’Ecole des Mines de Nantes.
Chargé de la rédaction d’un cahier des charges pour la création d’une plate-
forme logistique (voir section 6.2).
1996 – 1998 Université Pierre et Marie Curie, Paris
Doctorat de l’Université Pierre et Marie Curie (Paris VI).
Spécialité Informatique et Recherche Opérationnelle.
Laboratoire Ecole des Mines de Nantes,
d’accueil département Automatique et Productique.
Soutenance le 11 Décembre 1998 à l’université Pierre et Marie Curie
(Paris VI).
Sujet Etude de deux problèmes d’optimisation
en planification et ordonnancement.
Président du Jury P. Chrétienne (Professeur, Université P. et M. Curie, Pa-
ris).
Rapporteurs Y. Crama (Professeur, Université de Liège, Belgique),
J.B. Lasserre (Directeur de Recherche, LAAS/CNRS,
Toulouse)
Examinateurs M.-C. Portmann (Professeur, Ecole des Mines de
Nancy), C. Prins, (Maı̂tre Assistant, HDR, Ecole des
Mines de Nantes), S. Dauzère-Pérès (Maı̂tre Assistant,
HDR, Ecole des Mines de Nantes, Directeur de thèse).
Etude de deux problèmes d’optimisation en planification et ordonnance-

ment (problème de planification de la production en temps continu et pro-
blème général d’ordonnancement à une machine). Résolution par l’utilisa-
tion de techniques de recherche opérationnelle. Validation par le dévelop-
pement de logiciels prototypes.
10
1.4 Parcours et formation
1994 – 1995 Université Pierre et Marie Curie, Paris

Diplôme d’Etudes Approfondies de l’Université Pierre et Marie Curie (Pa-
ris VI).
Spécialité Informatique et Recherche Opérationnelle
Mention Bien
Responsable Ph. Chrétienne (Professeur, Université Pierre et Marie
Curie, Paris)
Sujet de DEA Les problèmes d’ordonnancement avec délais de com-
munication
Encadrement C. Picouleau (Maı̂tre de Conférences, CNAM, Paris).
1992 – 1994 Institut de Mathématiques Appliquées (IMA), Angers
Diplômes DEUG, Licence, Maı̂trise
obtenus
Spécialité Mathématiques Appliquées et Sciences Sociales
11
2 Enseignement
2 Enseignement
2.1 Enseignements à l’UVHC
Récapitulatif des heures enseignées à l’UVHC
Le tableau ci-dessous résume les heures d’enseignement depuis le recrute-
ment à l’université de Valenciennes et du Hainaut-Cambrésis. Le descriptif des
matières enseignées est présenté après le tableau. Dans ce tableau, sont reprises
les heures de cours magistraux (CM), de travaux dirigés (TD) et de travaux pra-
tiques (TP), ainsi que l’équivalence en heures TD (EqTD). L’IUT permet aussi de
comptabiliser certaines heures pour des tâches administratives (Responsabilité
des projets, Relations internationales), ainsi que des heures pour l’encadrement
des stages et des projets.
Année 2003-2004 Niveau CM TD TP EqTD

Recherche Opérationnelle IUT 2 et APPC 20h 2×40h – 110h
Informatique IUT 1 15h 2×20h 62.5h
Responsabilité de relations Internationales 10h 6.67h
Heures de stages et projets entrant dans le décompte du service 13h
Total année 2003-2004 192.17h

Informatique IUT 1 15h 2×14h 12h 58.5h
Mathématiques de la décision ISTV Master Info. 1 12h – – 18h
Responsabilité des projets 15h 10h
Total année 2002-2003 225.5h

Informatique IUT 1 15h 1×24h 20h 59.83h
Programmation linéaire EIGIP 2 – 1×18h – 18h
Responsabilité des projets 15h 10h
Heures de stages et projets entrant dans le décompte du service 19.5h
Total année 2001-2002 217.33h

Informatique IUT 1 – 1×25h – 25h
Programmation linéaire EIGIP 2 – 1×18h – 18h
Total année 2000-2001 220h

Qualité IUT 1 30h 1×30h – 75h
Recherche Opérationnelle IUT 2 – 1×40h – 40h
Informatique IUT 1 – 1×25h 2×25h 58.3h
Heures de stages et projets entrant dans le décompte du service 19.5h
Total année 1999-2000 192.83h
12
2.1 Enseignements à l’UVHC
Descriptif des interventions à l’UVHC

La ligne “# heures” correspond aux heures affectées à ce module pour les
étudiants. Pour les heures d’intervention me concernant, se reporter au tableau.
M17 – Algèbre et recherche opérationnelle
# heures CM 20h, TD 40h.
Public IUT OGP 2 et APPC
Lieu IUT OGP Cambrai
Année(s) depuis 1999
Responsabilité responsable du module depuis 2000
– Objectifs du cours : Donner à l’étudiant les principaux outils de la re-
cherche opérationnelle pour lui permettre d’optimiser les fonctions de
production et de logistique principales en entreprise. Acquérir les con-
nais sances de base en algèbre pour une poursuite d’étude éventuelle.
– Contenu du cours : Algèbre linéaire (bases de l’algèbre, inversion de ma-
trices) ; introduction à la recherche opérationnelle (historique, princi-
pales définitions et termes de la recherche opérationnelle) ; théorie des
graphes ; algorithmique des graphes (plus courts chemins, flot maxi-
mum, problèmes de transport) ; notions de programmation linéaire
(modélisa tion mathématique, propriétés et théorie de la programma-
tion linéaire) ; résolution graphique et algébrique ; applications indus-
trielles de la recherche opérationnelle (gestion de production, ordon-
nancement, MRP, logistique) ; cas particuliers de la programmation li-
néaire (matrice de contraintes totalement unimodulaires) ; langages de
modélisation ( apprentissage par l’exemple, principaux modèles ren-
contrés en industrie de production, résolution de cas industriels avec
Xpress-IVE de Dash Associates) ; mise en place d’un challenge depuis
2001 pour les élèves avec points de bonification pour les meilleurs
résultats obtenus.
– Notation : une note de TD reflétant le travail pratique en continu, un de-
voir maison, deux devoirs surveillés d’une heure, un devoir final de
deux heures.
M9 – Informatique
# heures CM 15h, TD 24h, TP 20h.
Public IUT OGP 1
Année(s) depuis 1999
Responsabilité responsable du module depuis 2001
– Objectifs du cours : Donner aux étudiants les moyens d’utiliser l’informa-
tique de manière avancée. Enseigner les bases de la programmation
structurée.
– Contenu du cours : Introduction à l’informatique et à la bureautique (trai-
tement de texte, tableur) ; introduction à la programmation structurée
13
2 Enseignement
(arbres programmatiques, langage structuré de 4e génération, concepts

de base, expressions, instructions, instructions conditionnelles, boucles,
fonctions, procédures) ; construction de macros évoluées en VBA ; con-
ception objet ; algorithmique générale ; mise en place de projets infor-
matiques (depuis 2001) pour la construction d’application VBA com-
plètes et utiles à un technicien OGP (un projet guidé, un projet libre) ;
liaison entre les programmes et les feuilles de calcul Excel (récupération
de données, écriture dans des feuilles de calcul, manipulation des ob-
jets Excel en VBA).
– Notation : deux devoirs surveillés d’une heure, un interrogation indivi-
duelle sur ordinateur, deux notes de projets.
MD – Mathématiques de la décision
# heures CM 12h, TD 18h, TP 6h.
Public Master Informatique 1
Lieu ISTV Valenciennes
Année(s) 2002-2003
Responsabilité responsable du module
– Objectifs du cours : Enseigner à l’étudiant les outils mathématiques néces-
saires à la résolution de problèmes dans le cadre de l’aide à la décision.
– Contenu du cours : Introduction générale à la programmation mathémati-
que (résolution linéaire, résolution entière, branch and bound en pro-
grammation entière, dualité) ; algorithmes de plus courts chemins ; al-
gorithmes de flots ; algorithmes de résolution des problèmes de trans-
ports ; modélisation mathématique ; mise en œuvre de la résolution
pratique en utilisant les logiciels professionnels (CPlex, MPL).
– Notation : deux devoirs surveillés, une note de TP.
F63 – Programmation linéaire
# heures CM 12h, TD 18h
Public EIGIP 2
Lieu Ecole d’ingénieurs de Valenciennes
Année(s) 2000-2002
Responsabilité intervenant de TD
– Objectifs du cours : Enseigner les bases de la programmation linéaire et de
la modélisation mathématique pour la résolution de problèmes d’opti-
misation industrielle.
– Contenu du cours : Modélisation mathématique (techniques de modélisa-
tion classiques, principaux modèles, modélisation entière ou binaire) ;
résolution graphique ; résolution algébrique (algorithme du simplexe,
dualité) ; analyse de sensibilité.
– Notation : un examen final.
14
2.2 Enseignements avant l’intégration à l’UVHC
M11 – Qualité
# heures CM 30h, TD 30h
Public IUT OGP 1
Année(s) 1999-2000
Responsabilité responsable du module
– Objectifs du cours : Donner aux étudiants les bases de la qualité en entre-
prise et les outils nécessaires à son application industrielle.
– Contenu du cours : Introduction à la qualité (historique, définitions) ; rap-
pels statistiques (moyenne, écart-type, variance, histogramme) ; démar-
che qualité en entreprise ; normes ISO ; outils mathématiques de la
qualité.
– Notation : deux devoirs d’une heure, un devoir final de deux heures, une
note de TD.
2.2 Enseignements avant l’intégration à l’UVHC

La formation pédagogique acquise depuis plusieurs années n’a pas débuté
à l’université de Valenciennes, mais remonte aux années 1992-1993. Depuis, un
certain nombre de matières ont été enseignées pour des public divers. La ligne “#
heures” correspond ici aux heures que j’ai effectué.
Programmation linéaire
# heures TD 35h, TP 52.5h
Public Elèves ingénieurs (Bac+3 et Bac +4)
Lieu Ecole des Mines de Nantes
Année(s) 1997-1999
Responsabilité chargé de travaux dirigés
Contenu Méthode du simplexe ; application à des problèmes
du monde in-
dustriel ; utilisa-
tion de logiciels
commerciaux.
Gestion de la production
# heures TD 27.5h
Public Elèves ingénieurs (Bac +4)
Lieu Ecole des Mines de Nantes
Année(s) 1997-1999
Contenu Gestion des stocks ; MRP ; planification des capacités.
15
2 Enseignement
Mathématiques
# heures 120h
Public CAP coiffure 1 et 2
Lieu Institut technique des études et des carrières à Angers
Année(s) 1993-1994
Responsabilité enseignant
Contenu Notions élémentaires de mathématiques ; fractions ;
pourcentages ; bases de la comptabilité.
Physique
# heures 60h
Année(s) 1993-1994
Contenu Notions de base en électricité ; dangers du courant
électrique ; applications pratiques en salon de coiffure.
Chimie
# heures 60h
Année(s) 1993-1994
Contenu Principes d’une réaction chimique ; application au traite-
ment des cheveux.
Informatique
# heures TD 60h
Public DEUG 1
Lieu Institut de mathématiques appliquées à Angers
Année(s) 1992-1993
Contenu Initiation à la programmation ; algorithmique.
2.3 Encadrements pédagogiques

Encadrements à l’UVHC
2003-2004
IUT-OGP 1 1 Projet : Développement des relations avec les anciens étudiants de
l’IUT-OGP, constitution d’un annuaire internet (8 étudiants).
IUT-OGP 2 1 Projet : Responsable du lancement du jeu d’entreprise FirStrat -
Stra&Logic (40 étudiants).
IUT-OGP 2 2 Stages : Sujets en cours.
16
2.3 Encadrements pédagogiques
2002-2003
IUT-OGP 1 1 Projet : Développement des relations avec les anciens étudiants de
l’IUT-OGP (8 étudiants).
IUT-OGP 2 2 Projets : Entreprise Blankaert Moto, assistance au choix d’une moto
(3 étudiants), UVHC, gestion des déchets (2 étudiants).
IUT-OGP 2 2 Stages : 1. et 2. Campack SA, organisation d’une ligne de produc-
tion de bouteilles de lait pendant sa mise en exploitation.
2001-2002
IUT-OGP 1 1 Projet : Réalisation d’un annuaire des anciens élèves de l’IUT OGP
de Cambrai (6 étudiants).
IUT-OGP 2 1 Projet : Entreprise Former - Réorganisation d’un atelier de frappe à
froid (6 étudiants).
IUT-OGP 2 3 Stages : 1. Babyliss SA, gestion des stocks pour le centre logistique,
2. Ygnis industrie, gestion de production, 3. SA Textiles Miersmann et Fils,
optimisation de production.
2000-2001
IUT-OGP APPC 2 Projets : 1. Société Amival, mise en place d’une ligne de pro-
duction alimentaire (4 étudiants), 2. Société Carolus Acier, aménagement
d’un atelier de fabrication (4 étudiants).
IUT-OGP 2 2 Stages : 1. Entreprise Bohain Textile, gestion de production, 2. Soci été
Sobotex, rédaction d’un cahier des charges pour un logiciel d’ordonnance-
ment.
1999-2000
IUT-OGP 2 3 Stages : 1. Usine d’embouteillage de Saint-Amand-les-eaux, normes
qualité, 2. Vieux-Condé estampages, gestion et suivi de production, 3. Mal-
lez imprimerie, suivi de production.
IUT-OGP 1 1 Projet : mise en place d’une démarche qualité au sein du départe-
ment OGP pour le suivi de la scolarité d’un élève (7 étudiants).
Encadrements avant l’intégration à l’UVHC

– Encadrement de mini-projets (découverte d’un outil spécifique de RO), élèves
ingénieurs de dernière année de l’Ecole des Mines de Nantes (1997-1999).
– Encadrement de stages à l’étranger des élèves de troisième année (niveau
maı̂trise) de l’Ecole des Mines de Nantes (1996-1999).
– Encadrement de projets transversaux en informatique sur le thème de la
planification et de l’ordonnancement d’ateliers (1998-1999).
17
2 Enseignement
2.4 Administration de l’enseignement

Elu au conseil restreint de l’IUT – depuis 2003
Le conseil restreint de l’IUT supervise l’attribution des postes d’enseignants-
chercheurs, d’enseignants et d’attachés temporaires d’enseignement et de recher-
che au sein de l’IUT tout entier. Cette attribution se fait sur proposition des dif-
férentes commissions de spécialistes et mon intervention concerne les métiers du
secteur secondaire.
Responsable des relations internationales – depuis 2003

La direction de l’IUT a décidé de motiver les départements pour relancer les
relations internationales. Trois axes sont privilégiés : 1) les échanges d’étudiants,
2) les échanges d’enseignants et 3) le développement d’axes de recherche en col-
laboration avec l’étranger. La mission du responsable des relations internatio-
nales est de promouvoir ces échanges et d’assister les différents intervenants du
département dans leurs démarches vers l’étranger en accord avec le directeur de
département et la direction de l’IUT.
Responsable de l’organisation des projets – 2001-2003

Axés sur la communication en première année à l’IUT, ces projets ont pour but
de faire travailler un groupe d’élèves autour d’un thème particulier et par la suite
de présenter le résultat de leurs travaux devant leurs collègues et professeurs. Ils
acquièrent ainsi une capacité à organiser le travail d’un groupe et présenter un
travail. Une seconde partie du projet consiste à proposer un produit innovant et
son processus de fabrication pour participer au concours national des OGP. Les
meilleures propositions seront étendues dans le cadre du projet de 2e année.
Les étudiants de deuxième année vont en début d’année réaliser un projet en
groupe au sein d’une entreprise. Ils seront mis face à des industriels à qui ils
doivent rendre des comptes. Ces projets portent principalement sur la rédaction
de procédures qualité et sur la réorganisation générale d’ateliers, le développe-
ment de modules informatiques, l’étude de logiciels de GPAO pour une implan-
tation sur site, etc. Depuis 2004, les projets en entreprise sont progressivement
remplacés par la participation au concours national OGP (création d’un produit
innovant et de son processus de production) et par la semaine du jeu d’entreprise.
Organisation des visites en entreprise – 1999-2001

Permettre aux élèves de connaı̂tre différents types d’entreprises et d’environ-
nements de travail est un des buts de ces visites. Un questionnaire étudié en fonc-
tion de chaque entreprise nous permet de voir comment les élèves appréhendent
leur futur cadre de travail industriel et éventuellement d’adapter certaines parties
de la formation. Il s’agit d’environ 4 à 5 visites par an.
18
2.4 Administration de l’enseignement
Recrutement, admission, jury

Participation aux jurys d’admission en DEA AISIH (2001-2003).
Participation au recrutement des élèves EIGIP (2000-2002).
Participation aux jurys d’attribution du DUT (depuis 2000).
Participation aux jurys de passage en IUT 2 (depuis 1999).
19
3 Supervision de travaux d’étudiants 2e et 3e cycle

3.1 Thèses de doctorat
Co-direction de Thèse européenne (2003-2006)
Frédéric BEUGNIES a débuté une thèse (bourse ministère) en septembre 2003

sur l’Optimisation multiobjectif de problèmes de routage dans les réseaux informatiques.
Encadrement : Dir. de thèse X. Gandibleux (UVHC), Co-encadrant M. Sevaux
(UVHC). Autres encadrants : S. RANDRIAMASY (Alcatel).
Frédéric BEUGNIES poursuit son travail de DEA sur l’étude de deux problèmes
d’optimisation dans le cas de l’exploitation de réseaux informatiques. Ces travaux
entrent dans le cadre d’une collaboration entre le LAMIH et ALCATEL.
Les questions posées relèvent de deux problèmes d’optimisation dans les graphes.
Dans un premier temps, il s’agit d’étudier des algorithmes dans un cadre temps-
réel et dans une seconde partie de faire de la gestion prévisionnelle de trafic. Bien
que ces deux approches portent sur des problèmes de caractéristiques distinctes
(la nature, la dimension, les contraintes, etc), elles se rejoignent dans le sens o ù
l’optimisation finale ne devra pas se faire sur un seul objectif, mais sur plusieurs
à la fois.
Co-direction de Thèse (2003-2006)
Karim BOUAMRANE a débuté en Mars 2003 une thèse sur le thème Système
d’aide à la régulation d’un réseau de transport bi-modal. Encadrement : Dir. de th èse :
Ch. Tahon (UVHC). Co-encadrant : M. Sevaux (UVHC).
Karim BOUAMRANE s’intègre dans le cadre d’un projet coopératif SART (Sys-
tème d’Aide à la Régulation de Transport, anciennement projet COMPACT, voir
section 6). Dans ce projet, un certain nombre de modules permettant la régulation
du transport sont déjà opérationnels et ont été développés par les différents parte-
naires du projet. Le travail dévolu lors de cette thèse concerne la proposition et la
mise en place opérationnelle du système d’aide proprement dit par le biais d’un
moniteur d’activation. Un simulateur est intégré à l’application pour simuler le
comportement du réseau de transport avant l’implantation du SART.
Co-direction de Thèse (2000-2004)
Yann LE QUÉRÉ a débuté en mai 2000 une thèse (en contrat Cifre avec la
SNCF) sur le thème de la planification réactive des tâches de maintenance pour les
TGV. Directeur de Thèse : Ch. Tahon. Co-encadrants : M. Sevaux (UVHC) et D.
Trentesaux (UVHC). Superviseurs industriels : G. Martin (SNCF) et E. Blervacque
(SNCF). Soutenance : 30 juin 2004. Jury : B. Grabot, M. Gourgand (Rapporteurs),
20
3.2 Mémoires de DEA
M.-J. Huguet, A. Thomas (Examinateurs).
Face à l’accroissement du trafic ferroviaire tant au niveau du fret que des pas-
sagers, il est important de diminuer les temps d’immobilisation lors des t âches
de maintenance des TGV. Pourtant les tâches de maintenance sont complexes et
souvent incertaines. L’étude s’intéresse alors à proposer une méthode consistant
à mesurer l’impact d’une structure de décision et des activités décisionnelles sur
un ordonnancement réactif.
Cette thèse a donné lieu à la publication d’un article en revue [9], une communica-
tion au GRP [45], une conférence internationale avec actes [20], une présentation
aux journées doctorales d’automatique [16] et à la soumission d’un autre article
pour publication [33]. Un contrat d’accompagnement de la thèse sur trois ans a
été signé avec la SNCF pour un montant de 160 kF.
3.2 Mémoires de DEA

DEA AISIH (2003-2004)
Changhong BIAN a réalisé un stage de DEA au sein de notre équipe. Le sujet
était l’étude d’un problème d’ordonnancement juste-à-temps sur une machine.
Les outils développés pour cette étude se basent sur des heuristiques pour la
création de solutions initiales dans des algorithmes génétiques qui sont implantés
par la suite. Les algorithmes génétiques s’appuient sur un codage binaire parti-
culièrement adapté à ce problème.
DEA AISIH (2003-2004)

Yan QIU a réalisé un stage de DEA au sein de notre équipe. Le sujet était
l’Application d’un algorithme génétique pour un problème de CAO (ce sujet a été co-
encadré avec Yves Mineur).
Dans ce travail, il s’agissait de mettre en œuvre un algorithme génétique permet-
tant de résoudre un problème d’optimisation continu rencontré dans le cadre de
la modélisation de formes géométriques en CAO. Un algorithme génétique déjà
opérationnel a été fourni comme base d’étude et le travail a consisté à proposer
des améliorations importantes.
DEA AISIH (2002-2003)

Johann SAINT MARS, étudiant en DEA a réalisé un stage au sein de notre
équipe. Le sujet était l’étude d’un problème d’ordonnancement à une machine avec
minimisation du retard pondéré.
Dans ce travail, une approche par programmation linéaire a été privilégiée. En
faisant appel aux outils commerciaux et aux modèles les plus récents pour ce
problème, l’étudiant a évalué les résultats trouvés et les a comparé aux meilleurs
résultats des heuristiques connues.
21
DEA AISIH (2000-2001)
Christophe TILLEUL, étudiant en DEA a réalisé un stage au sein de notre

équipe. Le sujet était l’étude d’un problème d’ordonnancement bicritère à une ma-
chine (ce sujet a été co-encadré avec X. Gandibleux).
Les problèmes d’ordonnancement sont le plus souvent très délicats à résoudre
lorsqu’ils sont N P-complets. Dans le cas bi-critère, la difficulté augmente encore.
L’objectif de ce travail a été de résoudre efficacement le problème d’ordonnance-
ment à une machine avec dates de disponibilité pour les critères de minimisation
du nombre de tâches en retard et le retard total de ces tâches.
3.3 Projets de DESS

DESS ICHM (2003-2004)
Chris BAUCHOT et Sébastien LEGENDRE ont réalisé au sein de notre équipe

un projet de DESS ICHM. Le sujet est la proposition d’une interface permettant la
manipulation et l’édition de graphes (ce sujet a été co-encadré avec P. Thomin).
Suite à la publication du livre “Algorithmes de graphes” [1], la construction d’une
interface graphique pour manipuler les graphes et exécuter les différents algo-
rithmes développés dans le livre est devenue indispensable. Une nouvelle version
de ce livre sera proposée en C++ prochainement. L’interface a été développée en
QT et est toujours opérationnelle.
DESS CCI (2002-2003)
Shiva ROUHOLAMINI et Rachid HARMAOUI, étudiants en DESS “Connais-

sances Complémentaires en Informatique” ont réalisé un projet au sein de notre
équipe. Le sujet traité était l’étude d’un problème d’ordonnancement juste à temps.
Dans ce stage, une approche par heuristiques constitue le cœur du travail de-
mandé. Les étudiants ont pu développer un certain nombres d’heuristiques et les
résultats ont été comparés aux heuristiques classiques connues pour ce problème.
DESS ICHM (1999-2000)
Laurent HULIN a effectué son projet dans notre équipe. Le sujet est la Proposi-
tion d’une interface permettant la manipulation de références bibliographiques (ce sujet
a été co-encadré avec P. Thomin).
Il avait pour but de développer une interface qui permette d’effectuer des re-
cherches de références bibliographiques à partir de fichiers au format BibTEX.
L’outil a été développé en php.
22
3.4 Projets IUP GEII
3.4 Projets IUP GEII

IUP GEII (2003-2004)
Yohan COLIN et Nicolas KOWALSKI ont effectué un projet dans notre équipe.
Le sujet est l’Optimisation de tournées de véhicules.
Le but du projet est d’utiliser les outils de la suite “Optimisation” d’ILOG pour
résoudre le problème de tournées de véhicules dans sa version académique. Une
comparaison a été menée entre les résultats du logiciel commercial et les meilleures
solutions connues de la littérature.
IUP GEII (2001-2002)

Sébastien LECHARDEUR et Johan SAINT MARS ont effectué un projet dans
notre équipe. Le sujet est Ordonnancement à une machine : une approche génétique (ce
sujet a été co-encadré avec P. Thomin).
Le critère à optimiser est la retard pondéré total dans le cas où les tâches sont
soumises à des dates de disponibilité. L’issue de ce projet a permis de présenter
une communication à IFORS’2002 [42].
IUP GEII (2001-2002)

Mathieu LEGEZYNSKI et Stéphane DEBACKER effectuent un projet dans
notre équipe. Le sujet est Ordonnancement à une machine : une approche par recherche
dispersée (ce sujet a été co-encadré avec P. Thomin).
Les deux binômes travaillent en parallèle sur un même problème d’ordonnance-
ment mais avec deux approches différentes. L’approche par recherche dispersée
est plus difficile à mettre en œuvre mais a donné de meilleurs résultats.
23
4 Administration et animation de la recherche

4.1 Animation de la recherche
Sur le plan international
Création et pilotage du groupe EU/ME (EURO Working Group – European chap-

ter on Metaheuristics). Groupe européen de recherche sur le thème des métaheu-
ristiques (créé à la suite d’une école d’hiver, avec K. Sörensen – Université d’An-
vers, Belgique et C. Wynants – Université libre de Bruxelles, Belgique). Ce groupe
compte aujourd’hui 750 membres de 65 nationalités différentes, avec un potentiel
encore élevé. Le but de ce groupe, créé en mars 2000, est de donner l’opportunité
aux chercheurs et industriels de se rencontrer et partager leur expérience de l’uti-
lisation et du développement des métaheuristiques. D’autres renseignements sur
le groupe se trouvent à l’adresse suivante :
http://webhost.ua.ac.be/eume/
Parmi les tâches effectuées dans ce cadre, on retrouve
– l’administration du groupe proprement dite (gestion des membres, organi-
sation d’assemblées générales, etc),
– la co-gestion du site web du groupe (mailing list, forum de discussion, in-
formations sur les conférences, etc),
– les relations avec EURO, l’organisme de tutelle (participation à des réunions
générales, rédaction de rapports d’activités),
– la gestion du budget,
– l’organisation d’une manifestation internationale annuelle (depuis 2001).
Au niveau du laboratoire
Création et animation (de 2000 à 2002) des séminaires du LAMIH permettant à

toutes les équipes de proposer des présentations d’intervenants locaux ou exté-
rieurs. Ces séminaires ont lieu sur un rythme régulier (2 fois par mois environ).
Ce travail a été réalisé en collaboration avec plusieurs membres d’autres équipes
du LAMIH. Depuis 2002, la direction du laboratoire a confié cette tâche à son
secrétariat.
Au niveau de l’équipe
Chargé de la collecte des informations (publications et activités de recherche et

d’encadrement) pour la mise à jour du rapport d’activité.
Mise en place et gestion des rapports de recherche internes à l’équipe. Ces rap-
ports permettent de diffuser l’information et prendre date pour une publication
qui est soumise à une revue mais pas encore acceptée.
24
4.2 Organisation de manifestations à Valenciennes
4.2 Organisation de manifestations à Valenciennes

Congrès International francophone PENTOM
Dates : 27-29 mars 2003
Organisateur : LAMIH/SP
Président du CO : O. Sénéchal
Thème : Performance et nouvelles technologies en maintenance
Participants : 115
Congrès International IFIP FEATS
Dates : 12-14 juin 2001
Président du CO : D. Deneux
Thème : Modélisation avancée par features dans la conception produit
Participants : 80
Réunion du groupe Bermudes
Dates : 2 février 2001
Organisateurs : X. Gandibleux et M. Sevaux
Thème : Ordonnancement dans les ateliers flexibles
Participants : 40
Journées GRP
Dates : 27-28 octobre 2000
Président du CO : D. Trentesaux
Thème : Productique
Participants : 80
4.3 Organisation de manifestations en dehors de Valenciennes

3e Joint meeting EU/ME
Dates : 18-19 décembre 2003
Lieu : Université d’Anvers (Belgique)
Organisateurs : EU/ME et l’université d’Anvers
Comité d’organisation : W. Dullaert, M. Sevaux, K. Sörensen, J. Springael
Thème : Applications réelles des métaheuristiques
Participants : 60
Site web : http ://webhost.ua.ac.be/eume/workshops/reallife.html
2e Joint meeting EU/ME
Dates : 4-5 novembre 2002
Lieu : Paris, Carré des sciences
Organisateurs : EU/ME et le groupe français PM2O (programmation ma-
thématique multi-objectif)
Comité d’organisation : X. Gandibleux, M. Sevaux, K. Sörensen, V. T’Kindt
Thème : Les métaheuristiques multi-objectifs
25
Participants : 60
Site web : http ://webhost.ua.ac.be/eume/workshops/momh.html
1er Joint meeting EU/ME
Dates : 28 novembre 2001
Lieu : City university, London (UK)
Organisateurs : EU/ME et UK Local search group
Comité d’organisation : C. Glass, C. Potts, M. Sevaux, K. Sörensen
Thème : Métaheuristiques
Participants : 30
Financement : obtenu pour 3 jeunes chercheurs auprès d’EURO
4.4 Organisation / président de sessions

Organisation d’un cluster de sessions
Conférence : INFORMS Denver Colorado 2004
Dates : 24-27 octobre 2004
Thème principal : métaheuristiques
Sessions : Tabu and scatter search, Genetic algorithms, Hybrid search.
Organisation de session invitée
Conférence : MIC’2001, Porto
Dates : 16-20 juillet 2001
Thème : ordonnancement et métaheuristiques.
Président de session
2004 : EURO XX (semi-plenary session)
2003 : MIC 2003, EMO 2003
2002 : PMS 2002, ROADEF 2002
2001 : MIC 2001, MOSIM 2001, ORBEL 2001.
26
5 Visibilité, rayonnement et autres activités
5.1 Collaborations
Université polytechnique de Hong-Kong
Correspondant : C. Oǧuz
Thème : Approches hybrides pour le flow-shop multi-stages
Visites : M. Sevaux à Hong-Kong (11-21 mai 2004 et 4-16 mars 2003). C.
Oǧuz à Valenciennes (22-29 mai 2004 et 12-27 juillet 2003)
Description : Cette collaboration a lieu dans le cadre d’un programme d’ac-
tion intégré (voir section 6.3). C. Oguz et moi-même travaillons sur des
problèmes d’ordonnancement de type flow-shop avec de nouvelles approches
intégrant les métaheuristiques et les techniques de type propagation de con-
traintes.
Université d’Anvers (Belgique)
Correspondant : K. Sörensen
Thème : Métaheuristiques ; Robustesse
Visites : En moyenne trois par an depuis 2000 ; accueil de K. Sörensen en
moyenne trois fois par an depuis 2001.
Description : Développement de nouvelles métaheuristiques hybrides [31]
et de méthodes pour l’ordonnancement robuste [18]. Rédaction de deux ar-
ticles en commun [7, 31]. Gestion du groupe EU/ME et mise en place d’un
nouveau projet [37]. Présentation lors d’un séminaire [56].
Université de Clermont-ferrand
Correspondants : G. Fleury et P. Lacomme
Thème : Ordonnancement stochastique
Description : Présentation en commun d’un article sur le problème de l’or-
donnancement stochastique [14].
Université de Technologie de Troyes et de Clermont-ferrand
Correspondants : C. Prins (UTT) et P. Lacomme (UBP)
Thème : Tournées sur arcs ; algorithmes de graphes
Visites : deux en 2002 et deux en 2003
Description : Rédaction en commun d’un article sur le problème des tournées
sur arcs multiobjectifs [17]. Rédaction d’un livre [1].
Ecole des Mines de Nantes
Correspondant : S. Dauzère-Pérès
Thème : Ordonnancement à machines parallèles et à une machine
Visites : M. Sevaux à Nantes (4-9 février 2002 et 17-21 avril 2001)
Description : Développement de nouveaux algorithmes pour la résolution
d’un problème à machines parallèles [19]. Poursuite des travaux engagés
précédemment et rédaction de plusieurs articles en commun [5, 8, 10, 35].
27
Ecole des Mines de Nantes et Univ. de Technologie de Troyes

Correspondants : C. Guéret (EMN) et C. Prins (UTT)
Thème : Programmation linéaire
Description : Collaboration et co-écriture du livre Programmation Linéaire et
de sa traduction [2, 3].
Massachusetts Institute of Technology
Correspondant : S.B. Gershwin
Thème : Planification de la production en temps continu ; détermination de
politiques de contrôle
Visites : M. Sevaux au MIT (15 avril-17 mai 1997)
Description : Rédaction d’un article en commun [10].
Institut de Math. Appliquées et Univ. de Technologie de Compiègne
Correspondants : P. Baptiste (UTC), L. Péridy (IMA), E. Pinson (IMA) et D.
Rivreau (IMA)
Thème : Ordonnancement à une machine
Description : Collaboration dans le cadre de l’étude des problèmes d’ordon-
nancement à une machine. Partage d’informations et de jeux de données.
Partage des résultats. Présentation conjointe des résultats lors d’une journée
du Gotha [59].
5.2 Appartenance à des sociétés et des groupes de recherche

Sociétés nationales et internationales
SOGESCI-B.V.W.B
Description : société belge de Recherche Opérationnelle
Adhésion : depuis 2001
Activités : participation aux manifestations nationales [46].
EU/ME
Description : EURO working group, European chapter on metaheuristics
Activités : administration du groupe (voir section 4.1), présentation lors des
manifestations [37, 38, 40].
ROADéF
Description : société française de Recherche Opérationnelle et d’Aide à la
Décision
Activités : participation et présentations [39, 43, 47, 48] lors des manifesta-
tions nationales et régionales, participation aux assemblées générales, ac-
tion auprès de la société pour les écoles d’été et d’hiver d’EURO.
28
5.2 Appartenance à des sociétés et des groupes de recherche
INFORMS
Description : Institute For Operations Research and the Management Science
Activités : Organisation d’un cluster au congrès INFORMS 2004. Présentation
des travaux au congrès INFORMS 1998 [49].
Groupes Nationaux
GDR MACS/STP
Description : Groupe de recherche du CNRS en Modélisation Analyse et
Conduite des Systèmes, pôle Sciences et Techniques de la Production
Adhésion : depuis 2003 (création du GDR en 2003)
Activités : Participation aux manifestations nationales. Présentation au thème
ORT [53].
Club EEA
Description : Club des enseignants et des chercheurs en électronique, élec-
trotechnique et automatique.
GRP
Description : Groupe de Recherche en Productique
Adhésion : de 2000 à 2003
Activités : Présentation de travaux aux manifestations nationales [41, 44,
45], participation à l’organisation d’une manifestation à Valenciennes.
PM2O
Description : Groupe de travail de la ROADEF en Programmation Mathé-
matique Multi-Objectif
Activités : participation aux réunions du groupe, organisation d’une mani-
festation conjointe avec EU/ME. Présentations [40, 54].
Bermudes
Description : Groupe de recherche en ordonnancement d’ateliers flexibles
Activités : participation et présentation [63] aux manifestations nationales,
organisation d’une réunion Bermudes à Valenciennes (avec X. Gandibleux).
GOThA
Description : Groupe de recherche en Ordonnancement Théorique et Ap-
pliqué
Activités : participation et présentation [59, 65, 66] à de nombreuses réunions
du groupe (en particulier dans le thème flexibilité), rédaction d’un chapitre
de livre [32].
29
Groupes Régionaux
GRAISyHM
Description : Groupe de Recherche en Automatique, Informatique des Sys-
tèmes Homme-Machine, thème commande et pilotage
Adhésion : depuis 2000 Activités : participation et présentation [62] aux
réunions du groupe.
5.3 Fonction d’édition

Editeur associé de IJCI
Depuis 2004, membre du comité éditorial de la revue International Journal of
Computational Intelligence. Le but de cette revue internationale est de développer
des systèmes qui reproduisent l’analyse, la résolution des problèmes et les fa-
cultés d’apprentissage du cerveau. Ces systèmes apportent les avantages de la
connaissance et de l’intelligence pour la résolution des problèmes complexes. Les
domaines couverts par le journal incluent : les outils intelligents g énériques (tech-
niques et algorithmes), des applications (utilisant des techniques intelligentes) et
des technologies intelligentes naissantes.
Editeur associé de IJSP

Depuis 2004, membre du comité éditorial de la revue International Journal of
Signal Processing. Le but de la revue IJSP est de prendre en compte tous les as-
pects de la théorie et de la pratique du traitement du signal. La politique de la
revue est une diffusion rapide des connaissances et des expériences auprès des
ingénieurs et des scientifiques travaillant dans la recherche, le développement ou
l’application pratique du traitement du signal.
Numéro Spécial EJOR

Suite au 3e Joint EU/ME workshop en collaboration avec l’université d’An-
vers, un numéro spécial de la revue European Journal of Operational Research est en
préparation. Ce numéro spécial porte sur le thème de l’application des métaheu-
ristiques. Date limite de soumission : 7 avril 2004, fin prévisionnelle du processus
de relecture : octobre 2004.
Edition d’un livre

Suite au 2e Joint EU/ME workshop “MOMH”, il a été décidé de publier dans
un livre, une sélection d’articles soumis à un processus de relecture de type revue
internationale. Cet ouvrage est publié comme “Lecture Notes in Economics and
Mathematical Systems” [4]. Taux de sélection : 36%
30
5.4 Evaluation de la recherche
5.4 Evaluation de la recherche

Rapporteur de revues internationales
Régulier pour les revues et chapitres de livres

– European Journal of Operational Research – EJOR
– Kluwer (chapitres de livre)
– Journal of Heuristics
– Annals of Operations Research
– IEEE System Man and Cybernetics – IEEE-SMC
– Engineering Applications of Artificial Intelligence – EAAI
Occasionnel pour les revues et chapitres de livres
– Quaterly Journal of the Belgian, French and Italian Operations Research
Societies – 4OR
– International Journal of Modelling and Simulation
– IEEE Transactions on Robotics and Automation – IEEE TRA
– Information Systems and Operational Research – INFOR
– Robotics and Computer Integrated Manufacturing
– IIE Transactions
– Decision Support Systems
– Rairo - Operations Research
– Journal Européen des Systèmes Automatisés – JESA
– Hermes (chapitre de livre)
Comités scientifiques
MIC’2005 Metaheuristics International Conference

Lieu : Vienne, Autriche
Dates : Eté 2005
Charge : Appel à communications, relecture d’articles soumis.
PENTOM’2005 PErformance et Nouvelles TechnOlogies en Maintenance
Lieu : Rabat, Maroc
Dates : Pâques 2005
FRANCORO 2004 Conférence Francophone Intl. de Recherche Opérationnelle
Lieu : Fribourg, Suisse
Dates : 18-21 août 2004
EURO’2004 EURO Conference
Lieu : Rhodes Island, Grèce
Charge : Appel à communications, organisation des “medals and awards”,
recherche de conférenciers invités.
31
ISS’2004 International Symposium on Scheduling

Lieu : Yumebutai, Japan
Dates : 24-26 Mai 2004
Lieu : Kyoto, Japon
Dates : 25-28 août 2003
ESI XXI EURO Summer Institute
Lieu : Neringa, Lithuania
Date : 25 juillet-7 août 2003
Charge : Choix de candidat français.
PENTOM’2003 PErformance et Nouvelles TechnOlogies en Maintenance
Lieu : Valenciennes
Dates : 27-29 mars 2003
Charge : Relecture d’articles soumis.
Lieu : Porto, Portugal
Jury de thèses
Comité d’accompagnement de la thèse de N. Souaı̈ (Faculté polytechnique de
Mons, Belgique), Problèmes de “crew scheduling” pour les campagnes aériennes,
soutenance prévue en 2006.
Examinateur de la thèse de K. Sörensen (Université d’Anvers, Belgique), Robust-
ness in optimization, soutenue le 13 juin 2003.
Examinateur de la thèse de W. Ramdane-Chrérif (Université de technologie de
Troyes), Problèmes de tournées de véhicules sur arcs soutenue le 12 d écembre
2002.
Commission de spécialistes
Membre élu titulaire de la commission de spécialistes en 61e section depuis 2004.
Membre élu suppléant de la commission de spécialistes en 61e section en 2003.
32
6 Contrats industriels, projets de recherche et finan-
cements
6.1 Contrats industriels
Contrat Alcatel - Anvers/Marcoussis
Contrat international – 12 mois – 2003-2004
Co-responsable scientifique, correspondant pour le LAMIH dans le cadre de
cette collaboration : M. Sevaux (LAMIH/SP) – 4 participants.
Une clause de confidentialité a été signée pour ce projet. Les activités menées
lors de cette collaboration et les montants importants engagés ne pourront être
dévoilés que lorsque le contrat sera terminé et la clause de confidentialité levée.
Par contre, au fur et à mesure de l’avancement certaines parties de l’activité pour-
ront être rendues publiques avec l’accord d’Alcatel. Elles seront alors mentionnées
dans ce paragraphe et dès que possible brevetées et publiées.
Thèse Cifre SNCF

Contrat régional – 24.2ke (160kF) – 36 mois – 2000/2003
Responsable scientifique : C. Tahon (LAMIH/SP) – 4 participants.
Convention CIFRE avec l’EIMM (Etablissement Industriel de Maintenance du
MAtériel) d’Hellemmes. Deux personnes de la SNCF sont impliquées dans le
suivi du projet. Une description plus détaillée du sujet de thèse se trouve dans
la section 3.1.
Centre Hospitalier de Valenciennes

Contrat local – 48ke – 36 mois – 2000/2003
Ce contrat entre notre équipe et le Centre Hospitalier de Valenciennes avait pour
objectif de réaliser une étude complémentaire du schéma directeur pour les flux
logistiques en vue de la construction d’un nouveau bâtiment. En fait, avant même
de prendre en compte la réalisation du nouveau bâtiment, nous avons dû tra-
vailler sur des modifications internes structurelles (construction d’une nouvelle
blanchisserie, déplacement des cuisines). Ce travail a débuté au début de l’année
2000 et pour une durée de 3 ans. Pour cette étude, la simulation a été l’outil de
travail privilégié.
33
6 Contrats, projets et financements
6.2 Projets de recherche

Projet Coopératif – SART (Phase 2, COMPACT)
Contrat régional – 139ke (Part labo 42.7ke) – 18 mois – 2003/2004
Responsable scientifique : S. Hayat (INRETS) – 30 participants.
Dans cette seconde phase, le problème de transport multi-modal est directement
abordé et des propositions de modélisation et de résolution sont faı̂tes pour pro-
poser un système d’aide à la régulation du trafic d’un réseau de transport multi-
modal (bimodal tramway/bus). Valenciennes et la Semurval, société exploitant
les transports en commun de Valenciennes sont les principaux acteurs de ce pro-
jet. La construction future du tramway fait que la réalisation de ce projet devient
indispensable pour nos partenaires. C’est dans ce cadre que s’inscrit en partie la
thèse en co-tutelle de Karim Bouamrane (voir section 3.1). Présentation lors d’un
séminaire [62].
Projet Coopératif – Phase 1, MOST/TACT

Contrat régional – 150ke (Part labo 27.4ke) – 18 mois – 2001/2002
Méthodologies pour l’Optimisation dans les Systèmes de Transport et de Télé-
communications / Technologies Avancées dans le domaine de la Communication
et des Transports Terrestres. Dans le cadre du CPER 2000-2006, le GRRT Nord-
Pas de Calais a obtenu le financement pour le projet coopératif. Il s’agit dans un
premier volet de ce projet de développer des outils en commun pour la résolution
de problèmes liés aux transports mais pas exclusivement. Présentation lors d’un
séminaire [62].
Projet S.M.I.L.E.
Projet interne EMN – 2 mois – 1999
Responsable scientifique : P. Dejax (EMN/AUTO) – 2 participants.
Rédaction du cahier des charges du projet de développement d’une plate forme
d’essais d’un Système de Modélisation et d’Intégration de la Logistique d’Entre-
prise. Ce projet vise à permettre aux entreprises le suivi de la chaı̂ne logistique
avec un module spécifique pour le suivi en temps réel des camions de collecte ou
de livraison ou l’adaptation aux perturbations du réseau routier est instantanée.
34
6.3 Financements obtenus
6.3 Financements obtenus

Financements obtenus pour déplacements
Dans le cadre du support d’activités scientifiques, le ministère des affaires
étrangères peut venir en aide aux laboratoires et aux chercheurs pour financer
des déplacements :
Aout 2003 Financement individuel (850e) pour la conférence MIC’2003 au Ja-
pon. Présentation d’un article [15].
Juillet 1999 Financement d’équipe (530e – P. Castagliola, C. Prins et M. Sevaux)
pour la conférence IEPM’99 en Ecosse. Présentation d’un article [25].
Programme d’action intégré - Hong-Kong

Projet international d’échange bilatéral – 8ke – 24 mois – 2003-2004
Responsable scientifique : M. Sevaux (LAMIH/SP) – 3 participants.
Ce programme d’échange avec Hong-Kong, développé de manière individuelle
au laboratoire est le début d’une collaboration avec Ceyda Oǧuz. Le thème re-
tenu est le développement de méthodes de résolution pour des problèmes d’or-
donnancement avec l’intégration des techniques de propagation de contraintes
dans les métaheuristiques ou dans la programmation linéaire. Ce programme est
financé par l’Egide sous la forme d’un programme d’action intégré Procore. Vi-
sites : M. Sevaux à Hong-Kong (11-21 mai 2004 et 4-16 mars 2003). C. Oǧuz à
Valenciennes (22-29 mai 2004 et 12-27 juillet 2003)
Projet national prospectif – Jemstic

Contrat national – 13ke – 12 mois – 2002
Responsable scientifique : D. Trentesaux (LAMIH/SP) – 6 participants.
Conception de systèmes coopératifs pour l’aide au pilotage distribué du cycle
de vie des systèmes de production de biens et de services. Il s’agit de savoir ici,
si l’action de coopération est bénéfique, toujours, parfois, ou peu souvent pour
l’ensemble des partenaires et aussi de mesurer l’impact de la coopération sur les
modes de fonctionnement de chacun des acteurs.
35
7 Thématiques de recherche
Les différents thèmes de recherche décrits ci-dessous sont le fruit de plusieurs
périodes de travail. On notera les périodes DEA (1994-1995), thèse (1996-1998),
après-thèse (1998-1999) et MDC (depuis 1999). Chaque thème abordé reprend dans
un cartouche la période de travail dans laquelle les recherches ont été effectuées et
les résultats obtenus en terme de publication et de visibilité du thème (chaque pu-
blication n’est citée que dans un seul cadre, même si certains sujets se recoupent).
La figure 1 reprend les thématiques de recherche menées depuis la thèse. Dans
cette même figure, on retrouve aussi les outils nécessaires à la résolution des
problèmes.
Introduction
Le contexte industriel de plus en plus compétitif oblige les entreprises à sans
cesse améliorer leur productivité tant au niveau des biens que des services. Dans
ce cadre, on retrouve classiquement les différents niveaux de décision stratégique,
tactique et opérationnel. Pour la production de biens, on s’intéresse à deux aspects
correspondants à deux niveaux de décision : la planification de production (ni-
veau tactique) et l’ordonnancement (niveau opérationnel). Pour les services, on
s’intéresse plus spécifiquement aux tournées de véhicules.
Le problème de planification est abordé dans un premier temps sous un angle
nouveau, mettant en avant la notion d’horizon continu (par opposition au clas-
sique horizon discrétisé). Basée sur cette modélisation, la détermination de poli-
tiques de contrôles est abordée avant de terminer par un domaine en plein essort,
la planification réactive.
Les problèmes d’ordonnancement, plus académiques dans leur présentation,
concernent l’objectif particulier de la minimisation du nombre de jobs en retard.
Ce type d’objectif est résolu pour les problèmes à une machine et à machines
parallèles avec des dates de disponibilité. Le nombre de jobs en retard peut être
pondéré ou non. Par la suite, on s’intéresse au problème à une machine, sans
doute le plus difficile à résoudre (toute relaxation donne aussi un problème N P-
difficile au sens fort). Il s’agit du problème de la minimisation du retard pondéré
total avec dates de disponibilité sur une machine unique. Enfin, sont abordés des
problèmes de robustesse indispensables pour les industriels soucieux d’appliquer
ces méthodes dans le cas opérationnel.
Les tournées de véhicules attirent l’attention des chercheurs depuis de nom-
breuses années et pourtant il reste encore des possibilités pour améliorer les mé-
thodes existantes et en proposer de nouvelles encore plus performantes. Les tour-
nées de véhicules sur arcs représentent un intérêt particulier car on retrouve ces
problèmes dans la réalité avec par exemple le salage des routes en cas de verglas,
l’entretien des bordures des routes, la distribution de courrier et la collecte des
ordures ménagères par exemple. On s’intéresse alors à l’optimisation bi-objectif
de ce problème. Le cas mono-objectif est aussi résolu de manière très efficace
36

2004

Robust VRP

Bi−objective CARP

CARP

Graphs algorithms

1−Machine TWT

Robust scheduling

2003

Metaheuristics

//−Machine Weighted # Late Jobs

Reactive Planning

2002

F IG . 1 – Problèmes et outils abordés

2001

Heuristics

Linear and integer programming

1−Machine Weighted # Late Jobs

2000

1999

1−Machine # Late Jobs

Control Policies

1998

Production Planning

1997

1996

37
donnant aujourd’hui sur des jeux de tests de la littérature les meilleurs résultats.
Quand le problème de tournées de véhicules est à résoudre non plus sur les arcs,
mais sur les nœuds du réseau, on s’intéresse alors à produire des solutions ro-
bustes.
7.1 Planification de la production
Planification de la production en temps continu
Période : 1996–1998.
Résultats : un article publié en revue [11],
une conférence avec actes [28],
une conférence sans actes [51] et
deux séminaires [64, 65].
La planification de production se résout habituellement en discrétisant un ho-

rizon de planification. Pourtant dans le cas de la production “gros volumes”, il
est souvent préférable de conserver l’horizon intact et de considérer un problème
de planification de la production en temps continu [11]. Dans ce cas, la demande (ha-
bituellement exprimée en quantités) est remplacée par des taux de demande et
la production est alors exprimée elle aussi sous forme de taux [67]. Les taux de
demande sont supposés connus et constants par morceaux. L’horizon de pla-
nification est séparé en périodes où la demande est constante, et les périodes
peuvent avoir des longueurs différentes. La capacité de production étant parfois
insuffisante la rupture est autorisée. Ainsi, dans d’autres périodes, il est parfois
nécessaire de produire plus que les demandes courantes. Il faut donc déterminer
combien produire en plus, et pendant combien de temps.
Pour réduire encore les coûts, on autorise les taux de production à changer à
d’autres instants que ceux où la demande change. Ces dates, ainsi que les fins de
périodes, seront appelées temps de changement [51]. Ainsi, il faut déterminer à
la fois les taux de production et les temps de changement. Parce que ce double
problème est trop difficile à résoudre simultanément, une procédure itérative à
deux étapes est proposée.
Dans un premier temps, avec des temps de changement fixés, les taux de
production optimaux sont déterminés à l’aide d’un modèle de programmation
linéaire. Dans une seconde étape, avec des taux de production fixes, les temps
de changement redondants sont éliminés, et de nouveaux instants sont ajoutés à
l’aide de règles. La procédure itère entre les deux étapes jusqu’à ce qu’il n’y ait
plus d’amélioration significative (voir [11, 64, 65, 67]).
Une extension permettant de traiter la demande discrète dans le cas où l’on
conserve le modèle en temps continu a aussi été développée [27, 67].
38
7.2 Ordonnancement
Détermination de politiques de contrôle

Période : 1997–1999.
Résultats : une conférence avec actes [25].
Dans le cas particulier où la demande est constante au cours du temps (sur
un intervalle de temps réduit par exemple), il peut être intéressant de déterminer,
partant d’un état des stocks quelconque, comment se ramener à un état station-
naire, i.e., déterminer une politique de contrôle. De tels résultats sont particu-
lièrement utiles pour un responsable de production dans le cadre de l’aide à la
décision. Pour le cas simple, réduit à une seule machine, une preuve d’optima-
lité de la politique développée est donnée [25]. Dès que plusieurs machines sont
mises en jeu, des conditions particulières sont indispensables pour que cette po-
litique reste optimale. Les précédents travaux ont permis de trouver des contre-
exemples dans le cas général [67]. L’utilisation des travaux de la planification en
temps continu a permis de valider les preuves et les contre-exemples.
Planification réactive
Période : 2000–2003.
Résultats : un article publié en revue [9],
deux conférences avec actes [16, 20] et
deux conférences sans actes [43, 45].
Dans le cadre de la thèse Cifre de Yann Le Quéré au sein de l’établissement in-
dustriel de la maintenance du matériel ferroviaire de la SNCF, l’étude s’intéresse à
proposer une méthode consistant à mesurer l’impact d’une structure de décision
et des activités décisionnelles sur un ordonnancement réactif. Un modèle per-
mettant d’intégrer les aléas possibles sur les tâches est développé. Il permet de
prendre en compte non-seulement ces aléas [20, 43, 45], mais aussi les délais
de communication entre les différents centres de décision [9]. Une approche par
réseaux de pétri [16] est associée à des techniques de propagation de contraintes.
7.2 Ordonnancement
Minimisation du nombre de jobs en retard sur une machine
Période : 1997–1999.
Résultats : un article accepté [5],
deux conférences avec actes [27, 29],
deux conférences sans actes [48, 49] et
un séminaire [66].
Sur l’ensemble des problèmes d’ordonnancement, les problèmes à une ma-
chine ont été les premiers étudiés, pour leur apparente simplicité. Malheureu-
sement, la plupart des problèmes généraux d’ordonnancement à une machine
39
sont N P-complets. Tout au long de cette étude, c’est le problème général d’or-
donnancement à une machine dans lequel l’objectif est de minimiser le nombre
de tâches en retard ou le nombre de tâches pondérées en retard, qui sera étudié
[67]. Pour ce problème, l’objectif était de proposer une méthode de résolution
exacte. A travers une nouvelle notion de séquence maı̂tre [67] des formulations
mathématiques ont été proposées [27] et ont permis de résoudre efficacement le
problème en dérivant des bornes inférieures par relaxation Lagrangienne [49]. Un
nouveau cas particulier polynomial [27] a été détecté et peut être résolu optimale-
ment en O(n log n). De ces résultats, une méthode arborescente efficace a permis
d’obtenir les meilleurs résultats au moment de leur publication [5, 29, 48].
Minimisation du nombre pondéré de jobs en retard sur une machine
Période : 1997–2002.
Résultats : deux articles publiés en revue [8, 10],
participation à une école d’hiver [24],
deux conférences sans actes [47, 50]
trois séminaires [59, 60, 63] et
un article en re-écriture [35].
L’étude du problème précédent apporte de nombreux éléments de réponse

pour le problème de la minimisation du nombre pondéré de tâches en retard pour
l’ordonnancement à une machine. Pour la plupart, les méthodes utilisées dans le
cas non-pondéré ont été étendues au cas pondéré [10, 26, 35, 50, 59]. L’ajout de
poids à chacun des jobs permet même d’améliorer certaines techniques et no-
tamment la résolution par relaxation Lagrangienne [10, 60, 63]. Les algorithmes
génétiques [8, 23, 24, 26] ont aussi permis à résoudre un certain nombre d’ins-
tances qui résistaient aux autres approches.
Minimisation du retard pondéré total sur une machine
Période : 2002–2003.
Résultats : une conférence sans actes [42].
En continuant à étudier les problèmes d’ordonnancement à une machine, on

s’intéresse désormais à la minimisation du retard pondéré total. Dans le cas gé-
néral (avec des dates de disponibilité des tâches différentes), le problème est
extrêmement complexe, puisque toutes les réductions simples conduisent à des
sous problèmes qui sont toujours N P-difficiles au sens fort. Une approche par
métaheuristiques donne des résultats parfois compétitifs par rapport aux meil-
leurs résultats connus [42].
40
7.2 Ordonnancement
Minimisation du nombre pondéré de jobs en retard sur machines parallèles

Période : 1999–2003.
Résultats : trois conférences avec actes [19, 21, 22],
une conférence sans actes [46],
deux séminaires [61, 62] et
un article en re-écriture [34].
Le critère de la minimisation du nombre pondéré de jobs en retard est main-
tenant étudié sur un environnement à machines parallèles. Ce problème devient
alors beaucoup plus compliqué que celui à une machine. Les réductions à des
sous problèmes voisins restent difficiles. On ne peut donc pas calculer facile-
ment des bornes sur le problème. Plusieurs solutions approchées sont données
par des heuristiques et des métaheuristiques [46, 61, 62]. Une approche basée sur
la méthode taboue [22] dans laquelle plusieurs raffinements sont proposés donne
des résultats très compétitifs. Une exploration plus large des métaheuristiques et
des heuristiques de ce problème [21] permet de donner de très bons résultats, jus-
qu’à 100 jobs et 6 machines [34]. Une relaxation Lagrangienne (celle du problème
à une machine) est étendue au cas machines parallèles et semble particulièrement
bien s’adapter [19] pour les problèmes de taille raisonnable.
Construction d’ordonnancements robustes

Période : 2001–2003.
Résultats : un article accepté en revue [7],
un séminaire [57] et
un article soumis [33].
Dans l’industrie automobile, la plupart des sous-traitants se sont install és à
proximité des usines d’assemblage. Ils délivrent les pièces prêtes à monter plu-
sieurs fois par jour selon les directives de l’usine d’assemblage (basées sur la
car sequence). Malheureusement, il arrive que les sous-traitants ne puissent li-
vrer à temps. On se propose alors de calculer un ordonnancement des pièces
à assembler sachant que les dates de début au plus tôt sont sujettes à modifi-
cation [18]. L’objectif ici est un indicateur de performance, la minimisation du
nombre pondéré de tâches en retard. Le problème étudié est 1|r j | ∑ w j Uj ou les
r j peuvent varier. Une approche par algorithme génétique [7] est proposée. A
chaque itération du GA, on évalue une série de fonctions dont les r j ont été mo-
difiés aléatoirement. Le résultat est un ordonnancement plus robuste qui peut
absorber des variations des r j . Ce type d’algorithmes génétiques modifiés permet
de proposer des solutions plus générales pour des problèmes où l’approche sta-
tistique n’est pas envisageable [33, 57]. Durant la thèse de Yann Le Quéré, une
problématique d’ordonnancement à contraintes de ressource ou les durées des
tâches sont variables a pu être résolue [14, 39].
41
Ordonnancement avec délais de communication

Période : 1994–1995.
Résultats : un rapport de DEA [68].
Le problème étudié est habituellement dénoté UET-UCT. Le but du stage de
DEA [68] était de donner des éléments de réponse (en terme d’optimisation) pour
le problème considéré. Dans un premier temps, le cas général a été présenté avec
une modélisation par programmation linéaire en nombres entiers. Des bornes
inférieures et supérieures ont été établies. Deux cas particuliers sont aussi présen-
tés : - le cas réduit à deux processeurs et - le cas où le graphe des précédences est
une arborescence. Par la suite, plusieurs algorithmes ont été établis et implémentés
en langage ADA pour résoudre le problème par heuristiques.
7.3 Tournées de véhicules

Optimisation multi-objectif en tournées de véhicules sur arcs
Période : 2001–2003.
Résultats : un article accepté en revue [6],
une conférence avec actes [17],
un séminaire [54].
La collecte des déchets en milieu urbain est une des nombreuses applications
du capacitated arc routing problem. Une approche multi-objectif est proposée pour
à la fois réduire la durée totale des tournées mais aussi équilibrer l’ensemble des
tournées [40, 54]. Ces travaux permettent de proposer des résultats sur les ins-
tances classiques [17] et ont prouvé leur intérêt [6]. Les problèmes multi-objectifs
sont bien souvent difficiles à résoudre et les métaheuristiques peuvent apporter
une réponse à ces problèmes [4].
Métaheuristiques avancées pour les tournées de véhicules sur arcs

Période : 2003–2004.
Résultats : une conférence avec actes [13] et
deux séminaires [52, 53].
Le problème de tournées de véhicules sur arcs dans sa forme académique est
déjà largement étudié [52]. Pourtant les derniers développements des algorithmes
à base de population ont permis de trouver encore des améliorations [53]. Les der-
niers résultats [13] sont aujourd’hui les plus performants. Ils sont même meilleurs
que les résultats précédemment publiés sur un ensemble de données tests de la
littérature à la fois en terme d’écart à la meilleure solution et en temps de calcul.
Dans ce cas précis, sur la base d’un algorithme génétique, une technique de ges-
tion fine de la population a été rajoutée dans l’algorithme et permet de gérer une
population plus petite et de meilleure qualité.
42
7.4 Autres approches ou problématiques
Robustesse en tournées de véhicules

Période : 2001–2003.
Résultats : une conférence avec actes [12].
Les problèmes de tournées de véhicules, largement étudiés dans la commu-
nauté scientifique, sont maintenant biens résolus par de nombreuses techniques.
Pourtant, toutes ces méthodes se basent sur une hypothèse très forte : la connais-
sance à priori exacte des données sur lesquelles s’appuie la problématique. Toutes
ces méthodes peuvent alors voir leurs performances s’effondrer si les données
varient, ne serait-ce qu’infimement. Dans cette étude, on s’intéresse à résoudre le
problème permettant de proposer des solutions qui resteraient satisfaisantes pour
le décideur même après variation. Encore une fois, ce sont les métaheuristiques
qui permettent de résoudre le problème en y incluant un cycle de simulation [12].

Optimisation en conception CAO
Période : 2003–2004.
Résultats : une conférence sans actes [38] et
un article soumis [30].
Encore une fois, les métaheuristiques peuvent être d’un grand secours, même
dans le domaine de l’optimisation continue. Dans le cas de la conception CAO
en automobile, il est indispensable de pouvoir reconstituer des courbes à partir
de séries de points mesurés sur un prototype. En définissant une fonction objectif
particulière et adaptée au problème considéré, on cherche à optimiser un certain
nombre de paramètres constituant les courbes modélisant la forme du prototype.
Le nombre de paramètres est important et nécessite un réglage minutieux difficile
à trouver manuellement. Un algorithme génétique “continu” permet de faire une
recherche efficace de ces paramètres et de trouver des configurations inattendues
et pertinentes. Ces résultats ont été présentés lors d’une conférence [38] et ont
donné lieu à la soumission d’un article en revue [30].
Optimisation multiobjectif
Période : 2003–2004.
Résultats : un livre édité [4].
La programmation mathématique multiobjectif est un vaste domaine cou-
vrant entre autre l’optimisation d’un problème donné en tenant compte de plu-
sieurs critères à la fois. Il ne s’agit pas de classer les critères entre eux, mais bien
au contraire de proposer des solutions de compromis lorsque les critères sont an-
tagonistes.
Avec l’aide du groupe français PM20, EU/ME a organisé en France deux jours
sur l’optimisation multiobjectif par métaheuristiques. Cette manifestation a réuni
43
60 chercheurs du monde entier. Lors de celle-ci, quatre sessions plénières ont per-
mis aux jeunes chercheurs de trouver ou retrouver un certain nombre d’informa-
tions indispensables pour leurs recherches. Une sélection d’articles soumis à un
processus de relecture de type revue internationale a été publié dans un “Lecture
Notes in Economics and Mathematical Systems” [4].
Métaheuristiques
Période : 1999–2004.
Résultats : une conférence avec actes [15],
deux conférences sans actes [41, 44] et
trois séminaires [55, 56, 58],
un article soumis [31] et
un chapitre en préparation [32].
Ce thème transversal est étudié à travers des problèmes d’optimisation et en

particulier des problèmes d’ordonnancement [41, 44]. La présentation de syn-
thèses importantes lors de séminaires invités [55, 56, 58] permet de mettre en
évidence l’information sur ces méthodes aujourd’hui incontournables pour la
résolution des problèmes d’optimisation. En s’appuyant sur l’expérience acquise
par l’étude de nombreux problèmes d’optimisation, on est à même aujourd’hui
de pouvoir proposer de nouvelles méthodes d’optimisation qui permettent de
palier à certaines des lacunes des métaheuristiques “classiques”. en particulier,
la gestion de la population est un point critique et ce dernier est partiellement
résolu dans une nouvelle méthode hybride appelée GA|PM [31]. De plus, pour
des problèmes de robustesse ou de flexibilité, les métaheuristiques offrent une
possibilité rapide et simple de résolution des problèmes. Un chapitre dans un
ouvrage collectif est en préparation sur ce thème [32].
Un livre de référence sur les métaheuristiques, cette fois-ci en Anglais, est en
préparation. Cet ouvrage va regrouper à la fois les développements théoriques
nécessaires à la compréhension des métaheuristiques et des programmes per-
mettant au lecteur de comprendre et d’appliquer rapidement les méthodes sur
ses propres problèmes. Les étudiants de troisième cycle et post-doctorants sont le
public de ce livre qui sera co-écrit avec C. Prins et K. Sörensen.
Période : 1996–2003.
Résultats : deux livres [2, 3].
La programmation linéaire est une branche de l’optimisation permettant de

résoudre de nombreux problèmes économiques et industriels. L’apparition de lo-
giciels puissants permet, aujourd’hui, de mettre cet outil à la disposition d’un
large public.
44
Ce livre, le premier du genre en français [3], met l’accent sur la démarche

de modélisation, appliquée ensuite à la résolution de 65 problèmes concrets re-
groupés par discipline en onze chapitres. À côté de problèmes industriels clas-
siques (comme le transport ou l’ordonnancement), vous y trouverez des champs
d’application moins connus ou plus récents, comme l’économie, la gestion du
personnel ou le secteur public.
Chaque chapitre est consacré à un domaine d’applications, et comporte six
problèmes concrets et suffisamment compliqués pour ne pas être traitables à la
main. À partir du descriptif de chaque cas, vous apprendrez, pas à pas, à construire
un modèle de programmation linéaire que vous pourrez résoudre et modifier
grâce au puissant logiciel Xpress fourni sur le CD-Rom d’accompagnement. Des
compléments à la fin de chaque chapitre et une bibliographie très complète per-
mettront aux curieux d’en savoir plus. Loin d’être un recueil d’exercices corrigés,
ce livre est en fait auto-contenu grâce à deux chapitres de cours sur les bases
théoriques de la programmation linéaire et à un chapitre sur les logiciels du
marché et les ressources web. Une version en Anglais de ce livre existe [2].
Algorithmes de graphes
Période : 2003–2004.
Résultats : un livre [1].
Les graphes et leurs algorithmes sont des outils mathématiques utilisés pour
modéliser et résoudre des problèmes complexes dans des domaines aussi variés
que l’optimisation (production industrielle, aide à la décision), la conception de
réseaux (électriques, routiers, télécoms) ou la modélisation de systèmes évolutifs
(économie, automatique). Cet ouvrage va permettre à tous de mettre en œuvre
des algorithmes plus ou moins complexes pour résoudre de tels problèmes. Ce
livre est une version complètement refondue de l’ouvrage ”Algorithmes de gra-
phes avec programmes en Pascal”, de Christian Prins, publié aux éditions Ey-
rolles en 1994, épuisé et non-réédité depuis plusieurs années.
L’objet de ce livre est de rendre ces techniques fondées sur la théorie des
graphes accessibles à des non-mathématiciens et de montrer comment les mettre
en œuvre dans des cas concrets. Une première partie introduit les notions d’op-
timisation combinatoire et de complexité des algorithmes, et donne un large pa-
norama des méthodes existantes, des plus classiques aux plus récentes (recuit si-
mulé, tabou). La seconde partie traite des différents problèmes de graphes : che-
mins optimaux, flots, tournées, coloration, etc. Les algorithmes, soigneusement
justifiés, sont accompagnés de programmes en pseudo-code et en langage Delphi
(Pascal objet), ainsi que d’exemples d’applications commentées [1].
Un nouveau livre sur ce même thème est en préparation mais cette fois-ci en
utilisant le langage C++ qui offre une plus grande souplesse dans la manipu-
lation des structures de données. Cet ouvrage sera écrit avec un nouvel auteur,
spécialiste en C++, Philippe Thomin.
45
Ontologies pour la recherche opérationnelle

Période : 2003–2004.
Résultats : une conférence sans actes [37].
Définie comme “la spéculation de l’être en tant qu’être”, l’ontologie peut être
plus simplement décrite comme une représentation formelle de parties impor-
tantes de notre monde. Avant tout considéré comme un concept philosophique,
les ontologies sont utilisées de manière croissante comme des outils pour partager
le savoir et l’information.
L’intérêt croissant pour le langage XML (eXtensible Markup Language) pour
décrire l’information d’un très grand nombre de domaines variés (mathématique,
chimie, physique, informatique, etc.) n’a fait que développer l’intérêt pour les
ontologies. En recherche opérationnelle, de nombreux jeux de données (ou ins-
tances) de problèmes standards sont couramment utilisés pour tester la perfor-
mance des algorithmes d’optimisation créés par les chercheurs. Une des plus
grandes collection de ces problèmes est la OR-Library. Mais cette collection n’est
pas unique. Ces instances sont des sources d’informations très fortement struc-
turées qui sont très souvent échangées entre les chercheurs. A notre avis, l’échange
de ces données pourrait être grandement facilité en appliquant certains des résul-
tats de la littérature sur les ontologies.
Le but de ce projet est de développer et d’implémenter une nouvelle version
de la OR-Library qui aurait les caractéristiques suivantes :
– Les jeux de données sont en format XML. Pour chaque problème, un schéma
XML du jeu de données correspondant est fourni. Les instances peuvent
être facilement validées et on y accède en utilisant des interpréteurs XML.
Ces interpréteurs XML sont disponibles partout et peuvent facilement être
inclus dans les programmes des algorithmes.
– Les instances peuvent être stockées dans la bibliothèque de façon décen-
tralisée. Chaque chercheur qui crée un nouveau jeu de données pour un
problème spécifique peut charger/stocker ces données. Les jeux de données
ainsi chargés peuvent être validés en les comparant au schéma XML du
problème.
– De nouveaux problèmes peuvent être proposés en soumettant un fichier
de schéma XML correspondant au nouveau problème. Pour une nouvelle
soumission, une demande de commentaires (RFC : request for comments)
est envoyé à la communauté. Quand le nouveau problème est accepté par
la communauté des chercheurs, le schéma XML et les données peuvent être
ajoutées à la bibliothèque.
– Les solutions et les valeurs des fonctions objectifs pour chaque instance
sont inclus dans la bibliothèque. Saisir ces données se fait aussi de manière
décentralisée. Avec peu d’efforts, les chercheurs peuvent partager les résul-
tats des développements récents avec l’ensemble de la communauté. Les
comparaisons entre les différentes techniques de résolution est ainsi facilité.
En atteignant nos objectifs, il sera facile de convaincre la communauté de ve-
46
nir ”alimenter” notre bibliothèque et d’éviter l’éparpillement des collections de

données comme on a pu le constater ces cinq dernières années. Une présentation
de ce nouveau domaine de recherche a été proposée lors du 3e EU/ME joint work-
shop, [37].
47
8 Liste des publications
*Attention les auteurs sont toujours classés par ordre alphabétique sauf si
l’implication d’un ou plusieurs d’entre eux est plus forte.
Livres (auteur)
[1] P. Lacomme, C. Prins, and M. Sevaux. Algorithmes de graphes. Eyrolles, 2003.
ISBN 2-212-11385-4, 425 pages (in French).
[2] C. Guéret, C. Prins, and M. Sevaux. Applications of optimisation with Xpress-

MP. Dash optimization, 2002. ISBN 0-9543503-0-8, Translated by S. Heipke.
[3] C. Guéret, C. Prins, and M. Sevaux. Programmation linéaire. Eyrolles, 2000.

Livres (éditeur)
[4] X. Gandibleux, M. Sevaux, K. Sörensen, and V. T’Kindt, editors. Meta-
heuristics for multi-objective optimisation, volume 535 of LNEMS. Springer,
2004. ISBN 3-540-20637-X, 230 pages.
Revues Internationales
[5] S. Dauzère-Pérès and M. Sevaux. An exact method to minimize the number
of tardy jobs in single machine scheduling. Journal of Scheduling, 2004. In
Press.
[6] P. Lacomme, C. Prins, and M. Sevaux. A genetic algorithm for a bi-objective

capacitated arc routing problem. Computers and Operations Research, 2004. In
Press.
[7] M. Sevaux and K. Sörensen. A genetic algorithm for robust schedules in a

just-in-time environment. 4OR – Quaterly journal of the Belgian, French and
Italian Operations Research Societies, 2(2) :129–147, 2004.
[8] M. Sevaux and S. Dauzère-Pérès. Genetic algorithms to minimize the weigh-

ted number of late jobs on a single machine. European Journal of Operational
Research, 151(2) :296– 306, 2003.
[9] Y. Le Quéré, M. Sevaux, C. Tahon, and D. Trentesaux. Reactive scheduling

of complex system maintenance in a cooperative environment with commu-
nication times. IEEE-SMC, 33(2) :225–234, 2003.
48
[10] S. Dauzère-Pérès and M. Sevaux. Using lagrangean relaxation to minimize
the weighted number of late jobs. Naval Research Logistics, 50(3) :273–288,
2003.
[11] S. Dauzère-Pérès, S.B. Gershwin, and M. Sevaux. Models and solving proce-
dures for continuous-time production planning. IIE Transactions, 32(2) :93–
103, 2000.
Conférences avec actes

[12] K. Sörensen and M. Sevaux. Robust and flexible vehicle routing in practical
situations. In Proceedings of 5th triennial symposium on transportation analysis,
TRISTAN V, Le Gosier, Guadeloupe, France, 13-18 June 2004. To appear.
[13] C. Prins, M. Sevaux, and K. Sörensen. A genetic algorithm with population

management (GA|PM) for the carp. In Proceedings of 5th trienal symposium
on transportation analysis, TRISTAN V, Le Gosier, Guadeloupe, France, 13-18
June 2004. To appear.
[14] G. Fleury, Lacomme P, and M. Sevaux. Stochastic maintenance scheduling

problem. In Proceedings of ninth international workshop on project management
and scheduling, PMS 2004, pages 405–409, Nancy, France, 26-28 April 2004.
[15] K. Sörensen and M. Sevaux. GA|PM : genetic algorithms with population

management for permutation problems. In Proceedings of 5th Metaheuristics
International Conference, MIC 2003, pages 69/1–7, Kyoto, Japan, 25-28 August
2003.
[16] Y. Le Quéré, M. Sevaux, C. Tahon, and D. Trentesaux. Modèle de coopération

d’un processus de ré-ordonnancement distribué. In Proceedings of the Auto-
matic control doctoral days (Journées doctorales d’automatique), pages 401–406,
LAMIH, Valenciennes, France, 25-27 June 2003. ISBN 2-905725-53-2.
[17] P. Lacomme, C. Prins, and M. Sevaux. Mutliple objective capacitated arc

routing problem. In Proceedings of 2nd International Conference on Evolutionary
Multi-Criterion Optimization, EMO’2003, pages 550–564, Faro, Portugal, 8-11
April 2003. LNCS 2632.
[18] M. Sevaux and K. Sörensen. Genetic algorithm for robust schedules. In

Proceedings of 8th International Workshop on Project Management and Scheduling,
PMS’2002, Valencia, Spain, April, 3-5 2002.
[19] S. Dauzère-Pérès and M. Sevaux. Lagrangean relaxation for minimizing the

weithed number of late jobs on parallel machines. In Proceedings of 8th Inter-
national Workshop on Project Management and Scheduling, PMS’2002, Valencia,
Spain, April, 3-5 2002.
49
[20] Y. Le Quéré, M. Sevaux, D. Trentesaux, and C. Tahon. Planification réactive
des opérations de maintien et d’actualisation réglementaire et technologique
des systèmes complexes. In Proceedings of the International Conference on Com-
puter aided Maintenance, Rabat, Morocco, November, 7-8 2001.
[21] M. Sevaux and P. Thomin. Heuristics and metaheuristics for parallel ma-
chine scheduling : a computational evaluation. In Proceedings of 4th Meta-
heuristics International Conference, MIC 2001, pages 411–415, Porto, Portugal,
July, 16-20 2001.
[22] M. Sevaux and P. Thomin. Recherche taboue améliorée pour l’ordonnan-

cement sur machines parallèles. In Actes de la 3ième conférence Francophone
de MOdélisation et de SIMulation, MOSIM’01, pages 829–833, Troyes, France,
April, 25-27 2001.
[23] M. Sevaux and S. Dauzère-Pérès. A genetic algorithm to minimize the

weighted number of late jobs on a single machine. In Proceedings of 7th In-
ternational Workshop on Project Management and Scheduling, PMS’2000, pages
242–244, Osnabrük, Germany, April, 17-19 2000.
[24] M. Sevaux and S. Dauzère-Pérès. Building a genetic algorithm for a single

machine scheduling problem. In Proceedings of the 18th EURO Winter Insti-
tute, ESWI XVIII, Lac Noir, Switzerland, March, 4-18 2000.

the (weighted) number of late jobs. In National contribution for the 15th triennal
conference, IFORS’99, Beijin, P.R. of China, August, 16-20 1999.
[26] S. Dauzère-Pérès and M. Sevaux. Control policies for some continuous-time

production scheduling problems. In Proceedings of the International Conference
on Industrial Engineering and Production Management, IEPM’99.
[27] S. Dauzère-Pérès and M. Sevaux. Handling discrete demand in continuous-

time production planning. In Proceedings of the 9th symposium of the Interna-
tional Federation of Automatic Control on Information Control in Manufacturing,
INCOM’98, pages 467–472.
[28] S. Dauzère-Pérès and M. Sevaux. Various mathematical programming for-

mulations for a general one machine sequencing problem. In Proceedings
of the 4ème Journées Nationales sur la Résolution Pratique des Problèmes NP-
Complets, JNPC’98, pages 63–68.
[29] S. Dauzère-Pérès and M. Sevaux. A branch and bound method to mini-

mize the number of late jobs on a single machine. In Proceedings of the 6th
International Workshop on Project Management and Scheduling, PMS’98, pages
193–196.
50
Articles en cours de révision, soumis ou en re- écriture
[30] M. Sevaux and Y. Mineur. Continuous optimisation by GA for a curve fitting
problem in autmobile industry. Technical report, University of Valenciennes,
CNRS UMR 8530, LAMIH/SP, 2004. Submitted.

management. Technical report, University of Valenciennes, CNRS UMR
8530, LAMIH/SP, 2004. Submitted.
[32] M. Sevaux, K. Sörensen, and Y. Le Quéré. Flexibilité et robustesse en ordonnan-

cement, chapitre Métaheuristiques pour la planification et l’ordonnancement
robuste, pages 91–110. Hermes, 2004. Under revision.
[33] M. Sevaux and Y. Le Quéré. Solving a robust maintenance scheduling pro-

blem at the F rench railway company. Technical Report LAMIH/SP-2003-3,
University of Valenciennes, CNRS UMR 8530, LAMIH/SP, April 2003. Un-
der revision.
[34] M. Sevaux and P. Thomin. Heuristics and metaheursitics for a parallel

machine scheduling problem : a computational evaluation. Technical Re-
port LAMIH/SP-2001-2, University of Valenciennes, CNRS UMR 8530, LA-
MIH/SP, November 2001. Under revision.
[35] S. Dauzère-Pérès and M. Sevaux. An efficient formulation for minimizing

the number of late jobs in single-machine scheduling. Technical Report
98/9/AUTO, Ecole des Mines de Nantes, July 1998. Under revision.
Conférences sans actes

[36] C. Prins, M. Sevaux, and K. Sörensen. Genetic algorithms with population
management : an application to arc routing. In Optimization, Lisboa, Portu-
gal, 27-27 July 2004.
[37] K. Ven, K. Sörensen, M. Sevaux and J. Vereist. Lib-OR – Library of OR data

sets. In 3rd Joint EU/ME workshop with the university of Antwerp on Real-life
application of metaheuritics, Antwerp, Belgium, Dec. 18-19 2003.
[38] Y. Mineur and M. Sevaux. Curve fitting for styling application by genetic
algorithm. In 3rd Joint EU/ME workshop with the university of Antwerp on Real-
life application of metaheuritics, Antwerp, Belgium, Dec. 18-19 2003.
[39] Y. Le Quéré and M. Sevaux. Approche robuste pour un problème à

contraintes de ressources. In 5ième conférence nationale de la société française
de recherche opérationnelle, ROADEF’2003, Feb. 26-28 2003.
51
[40] P. Lacomme, C. Prins, and M. Sevaux. Multiobjective capacitated arc routing
problem. In 2nd Joint EU/ME workshop with the french PM2O group ( MOMH
Workshop) on multi-objective metaheuristics, Paris, France, Nov. 4-5 2002.
[41] M. Sevaux. Les méthodes de recherche à population. In Groupe de Recherche

en Productique - Organisation et Gestion de Production, Oct. 24-25 2002.
[42] M. Sevaux and P. Thomin. Scatter search and ga : a one machine scheduling
problem comparison. In The sixteenth triennial conference of international fe-
deration of operational research societies, IFORS’2002, Edinburgh, UK, Jul. 6-13
2002.
[43] Y. Le Quéré, M. Sevaux, D. Trentesaux, and C. Tahon. Résolution d’un

problème industriel de maintenance des TGV à la SNCF. In 4ième conférence
nationale de la société française de recherche opérationnelle, ROADEF’2002, Paris,
France, Feb. 20-22 2002.
[44] M. Sevaux. Les méthodes de recherche à voisinage. In Groupe de Recherche en

Productique - Organisation et Gestion de Production, Nov. 8-9 2001.
[45] Y. Le Quéré, D. Trentesaux, M. Sevaux, and C. Tahon. Gestion réactive des

opérations de maintien et d’actualisation réglementaire et technologiques
des systèmes complexes. In Groupe de Recherche en Productique – Organisation
et Gesion de Production, Montpellier, France, Mar. 29-30 2001.
[46] M. Sevaux and P. Thomin. Efficient heuristic and tabu search for parallel
machine scheduling. In ORBEL Conference, Antwerp, Belgium, Jan. 29-30
2001.
[47] M. Sevaux and S. Dauzère-Pérès. Un algorithme génétique pour minimiser

le nombre pondéré de jobs en retard sur une machine. In 3ième conférence na-
tionale de la société française de recherche opérationnelle, ROADEF’2000, Nantes,
France, Jan. 26-28 2000.
[48] S. Dauzère-Pérès and M. Sevaux. Une méthode arborescente améliorée pour

la minimisation du nombre de jobs en retard. In 2ième conférence nationale de
la société française de recherche opérationnelle, ROADEF’2000, Autrans, France,
Jan. 13-15 2000.
[49] S. Dauzère-Pérès and M. Sevaux. On minimizing late jobs in single machine

scheduling. In INFORMS/Cors’98, Montreal, Canada, Apr. 26-29 1998.

the weighted number of late jobs in one machine sequencing with release
dates. In Symposium on Combinatorial Optimization, CO’98, Brussels, Belgium,
Apr. 15-17 1998.
52
[51] S. Dauzère-Pérès, S.B. Gershwin, and M. Sevaux. Modeling and solving pro-
cedures of continuous-time production planning problems. In 8th Annual
Meeting of the Production and Operations Management Society, POM-97, Miami
Beach, Florida, USA, Apr. 12-15 1997.
Séminaires
[52] M. Sevaux. A genetic algorithm with population management GA|PM for
the CARP. Invited seminar, The Hong-Kong Polytechnic University, Depart-
ment of Logistics, 14 May 2004.
[53] M. Sevaux. Production de biens et services : application en ordonnancement
et transport. Séminaire invité, GDR MACS/STP groupe ORT (in French), 13
February 2004.
[54] M. Sevaux. Advances in mutliple objective capacitated arc routing problem.
Invited seminar, University of Valenciennes, French research group on mul-
tiple objective mathematical programming, 15 May 2003.
[55] M. Sevaux. Population metaheuristics. Invited seminar, The Hong-Kong
Polytechnic University, Department of Management, 14 March 2003.
[56] M. Sevaux. Metaheuristics : a quick overview. Invited seminar, University
of Antwerp, Faculty of Applied Economic Sciences, 7 February 2003.
[57] M. Sevaux. Métaheuristiques pour la résolution robuste de probl èmes d’op-
timisation combinatoire. Séminaire invité, Mathématiques Appliquées de
Bordeaux (in French), 14 November 2002.
[58] M. Sevaux. Les méthodes de recherche à voisinage. Séminaire invité, Faculté
Polytechnique de Mons, groupe Image (in French), 6 March 2002.
[59] M. Sevaux and D. Rivreau. Single machine scheduling : minimizing the
[weighted] number of late jobs, a review. Séminaire invité, Groupe de Re-
cherche en Ordonnancement Théorique et Appliqué, GOThA (in French), 28
September 2001.
[60] M. Sevaux. Single machine scheduling : Minimizing the [weighted] num-
ber of late jobs. Invited seminar, Université Libre de Bruxelles - Service des
Mathématiques de la Gestion, 10 May 2001.
[61] M. Sevaux and Ph. Thomin. Parallel machine scheduling : a metaheuristic
computational evaluation. Séminaire invité, Institut de Recherche en Com-
munication et Cybernétique de Nantes, IRCCyN (in French), 19 April 2001.
[62] M. Sevaux and P. Thomin. Parallel machine scheduling : a (meta)heuristic
computational evaluation. Séminaire invité, Groupe MOST (in French), 5
April 2001.
53
[63] M. Sevaux. Ordonnancement à une machine par relaxation lagrangienne.
Séminaire invité, Groupe de Recherche Bermudes (in French), 11 June 1999.
[64] S. Dauzère-Pérès and M. Sevaux. Modeling and solving procedures of

continuous-time production planning problems. Invited seminar, Massa-
chusetts Institute of Technology, Cambridge, MA, USA, 23 April 1997.
[65] S. Dauzère-Pérès and M. Sevaux. Modèles et procédures de résolution pour

la planification en temps continu. Séminaire invité, Groupe de Recherche en
Ordonnancement Théorique et Appliqué, GOThA (in French), 19 December
1997.
[66] S. Dauzère-Pérès and M. Sevaux. Minimisation du nombre de jobs dans le

problème d’ordonnancement à une machine. Séminaire invité, Groupe de
Recherche en Ordonnancement Théorique et Appliqué, GOThA (in French),
19 December 1997.
Thèse de doctorat
[67] M. Sevaux. Etude de deux problèmes d’optimisation en planification et ordonnan-
cement (On Two Optimization Problems in Production Planning and Scheduling).
Thèse de doctorat, Université Pierre et Marie Curie (Jussieu, Paris VI), 1998.
Ecole des Mines de Nantes.
Rapport de DEA
[68] M. Sevaux. Ordonnancement avec délais de communication. Rapport de
DEA, Université Pierre et Marie Curie (Jussieu, Paris VI), 1995.
54
Deuxième partie
Synthèse scientifique
56
1 Introduction générale
Présenter les différents problèmes abordés aux cours des années précédentes
n’aurait qu’un intérêt limité pour une telle synthèse. Dans ce document, nous
avons volontairement choisi de présenter les métaheuristiques utilisés pour la
résolution de nombreux problèmes d’optimisation. Ces outils seront mis en avant
et rassemblés par catégories. Mais avant tout, cette synthèse a pour modeste am-
bition de proposer un panorama général des plus importantes métaheuristiques
d’aujourd’hui qui permettent de résoudre un problème d’optimisation combina-
toire. La conclusion proposera ce que pourraient être, selon nous, les caractéristi-
ques d’une bonne métaheuristique.
1.1 Pourquoi les métaheuristiques ?
Depuis toujours, les chercheurs ont tenté de résoudre les problèmes N P-diffi-
ciles le plus efficacement possible. Pendant longtemps, la recherche s’est orientée
vers la proposition d’algorithmes exacts pour des cas particuliers polynomiaux.
Ensuite, l’apparition des heuristiques a permis de trouver des solutions en g énéral
de bonne qualité pour résoudre les problèmes. En même temps, les méthodes de
type “séparation et évaluation” ont aidé à résoudre des problèmes de manière
optimale, mais souvent pour des instances de petite taille.
Lorsque les premières métaheuristiques apparaissent, beaucoup de chercheurs
se sont lancés dans l’utilisation de ces méthodes. Cela a conduit à une avancée im-
portante pour la résolution pratique de nombreux problèmes. Cela a aussi créé un
engouement pour le développement même de ces méthodes. Il existe des équipes
entières qui ne travaillent qu’au développement des métaheuristiques.
Il faut aussi reconnaı̂tre que c’est un formidable outil pour la résolution effi-
cace des problèmes posés.
1.2 Intensification et diversification
Toutes les métaheuristiques s’appuient sur un équilibre entre l’intensification

de la recherche et la diversification de celle-ci. D’un côté, l’intensification permet
de rechercher des solutions de plus grande qualité en s’appuyant sur les solutions
déjà trouvées et de l’autre, la diversification met en place des stratégies qui per-
mettent d’explorer un plus grand espace de solutions et d’échapper à des minima
locaux.
Ne pas préserver cet équilibre conduit à une convergence trop rapide vers
des minima locaux (manque de diversification) ou à une exploration trop longue
(manque d’intensification). Cette vision est bien sûr simpliste mais malheureuse-
ment on l’observe trop souvent encore aujourd’hui. Il existe de nombreux articles
57
1 Introduction générale
soumis à publication qui se trouvent rejetés par les arbitres pour ces raisons.
Dans les sections qui vont suivre, nous avons essayé, de présenter chaque
métaheuristique de façon uniforme. D’abord, nous rappelons, quand c’est pos-
sible, les papiers où la méthode a été introduite la première fois. Une descrip-
tion textuelle de l’algorithme est proposé avec ses particularités. Parfois, il existe
des articles plus récents que les articles d’origine qui présentent soit une vue
synthétique soit un tutorial détaillé et que nous citons. Ensuite, nous donnons
une présentation algorithmique de chaque métaheuristique, avant de montrer
quels sont les facteurs d’intensification et de diversification de chacune de ces
méthodes.
1.3 Techniques de résolution pratique
Nous nous plaçons ici dans le contexte particulier d’un chercheur faisant face
à un problème d’optimisation pour lequel les méthodes exactes ont atteint leurs
limites. Nous introduisons alors un ensemble d’outils qui permettront de trouver
des solutions et de les améliorer. Pour chacune des sections ci-dessous, un cer-
tain nombre de méthodes sont présentées. Il est impensable de déconnecter les
méthodes des applications pour lesquelles elles ont été dessinées. Chacune des
sections propose donc une partie “Applications” mettant en œuvre les méthodes
proposées sur des problèmes que nous avons rencontrés pendant nos recherches.
Dans cette partie application, nous présenterons à la fois quelques références im-
portantes de la littérature et nos propres travaux.
Face à un problème d’optimisation, l’étape la plus importante est sans doute
l’étape de la modélisation (c’est à dire passer d’un problème réel à une formu-
lation mathématique de ce problème). C’est aussi l’étape la plus difficile à en-
seigner car la plupart du temps, c’est l’expérience qui permet de définir un bon
modèle. Dans notre cas, une fois ce modèle établit (et la preuve que le problème
appartient à la classe des problèmes N P-difficiles obtenue), nous essayons de
le transformer en un problème de programmation mathématique que nous al-
lons tenter de résoudre en utilisant un logiciel commercial (pour des exemples de
telles modélisations et résolutions, voir Guéret et al. [53, 54]). Utiliser la program-
mation mathématique permet aussi d’obtenir des bornes sur le problème et sans
doute de mieux appréhender sa difficulté propre.
Une fois obtenues des bornes et éventuellement quelques solutions par heuris-
tique , on peut commencer à mettre en œuvre des métaheuristiques. Pour améliorer
des solutions initiales, on peut utiliser des techniques de recherche locale. Les
méthodes de recherche locale, des plus simples aux plus compliquées sont un
atout majeur et certaines d’entre elles sont décrites dans la section 2. On y trou-
vera les méthodes de descente, de recuit simulé, de recherche tabou, de recherche
à voisinage variable, de GRASP, d’iterated local search, de guided local search.
58
Dans la section 3, les méthodes à base de population de solutions les plus
simples sont détaillées. Elles présentent un avantage indéniable, mettre en place
un parallélisme intrinsèque et explorer un espace de solutions très vaste. Les al-
gorithmes génétiques et les colonies de fourmis sont présentés dans cette sec-
tion. Pour diverses raisons, nous n’avons pas eu l’occasion de tester les colonies
de fourmis, seule une application en ordonnancement utilisant des algorithmes
génétiques sera détaillée.
Ces méthodes à population, dans leur version de base, ont souvent du mal à
trouver des solutions compétitives avec d’autres techniques. La section 4 présente
des améliorations indispensables à la réussite de telles approches. En ajoutant des
techniques de recherche locale aux algorithmes génétiques, on obtient les algo-
rithmes mémétiques. A côté, le scatter search, technique innovante, sera présenté
et testé. Le dernier point concerne un nouvel algorithme (GA|PM) présentant des
caractéristiques très intéressantes pour l’optimisation.
La section 5 présente des compléments intéressants. Il s’agit soit de travaux
supplémentaires que nous avons mené sur les métaheuristiques, soit des utili-
sations des métaheuristiques pour d’autres propos. Des techniques de réglages
automatiques de paramètres sont proposées, et des résolutions en robustesse, op-
timisation multiobjectif et optimisation continue sont présentées.
Dans la conclusion (section 6), nous mettons en avant, selon nous, les bonnes
caractéristiques que devrait posséder une métaheuristique aujourd’hui et surtout
quelques mises en garde importantes pour la conception de ces méthodes.
La section 7 expose les champs d’application qui nous intéressent particuliè-
rement aujourd’hui et que nous souhaitons traiter dans un futur proche. En pa-
rallèle à ces applications, il nous apparaı̂t indispensable de pouvoir proposer des
thèmes de recherche sur le plus long terme. D’un côté, une collaboration indus-
trielle pourrait déboucher rapidement sur une thèse dans le domaine de l’optimi-
sation du trafic dans les réseaux de télécommunications. De plus, nous souhaite-
rions voir avancer un travail de grande envergure sur la proposition d’une vue
unifiée des métaheuristiques à population.
2 Méthodes de recherche locale
Les méthodes de recherche locale ou métaheuristiques à base de voisinages

s’appuient toutes sur un même principe. A partir d’une solution unique x 0 , con-
sidérée comme point de départ (et calculée par exemple par une heuristique cons-
tructive), la recherche consiste à passer d’une solution à une solution voisine par
déplacements successifs. L’ensemble des solutions que l’on peut atteindre à par-
tir d’une solution x est appelé voisinage N (x) de cette solution. Déterminer une
solution voisine de x dépend bien entendu du problème traité.
59
f (x)
local minimum
global minimum
F IG . 2 – Minimum local et global
De manière générale, les opérateurs de recherche locale s’arrêtent quand une

solution localement optimale est trouvée, c’est à dire quand il n’existe pas de
meilleure solution dans le voisinage. Mais accepter uniquement ce type de solu-
tion n’est bien sûr pas satisfaisant. C’est le cas des méthodes de descente présen-
tées ci-dessous (voir section 2.1)
Dans un cadre plus général, il serait alors intéressant de pouvoir s’échapper de
ces minima locaux (voir Fig. 2). Il faut alors permettre à l’opérateur de recherche
locale de faire des mouvements pour lesquels la nouvelle solution retenue sera de
qualité moindre que la précédente. C’est le cas immédiat des méthodes de recuit
simulé (section 2.2), de la recherche tabou (section 2.3) et de l’iterated local search
(section 2.6). Dans le cas de la VNS, de GRASP et du guided local search, les tech-
niques pour échapper aux minima locaux sont différentes et seront expliquées
dans les sections 2.4, 2.5 et 2.7. Les méthodes de recuit simulé et tabou sont plus
anciennes et sans doute plus populaires. L’attention qui leur sera portée sera alors
un peu plus grande.
Certaines des méthodes détaillées ci-dessous ont fait l’objet de présentation
en conférences ou séminaires. Citons par exemple [98, 100, 101] qui présentent
à la fois des méthodes de descente, de recherche tabou et de recuit simulé. Bien
sûr, ces méthodes sont rarement présentées sans contexte applicatif. La section 2.8
présentera pour un problème d’ordonnancement à machines parallèles l’applica-
tion de quelques unes de ces méthodes. Pour une explication générale sur ces
méthodes de recherche, on peut se référer à l’ouvrage ancien mais assez complet
de Papadimitriou et Steiglitz [88].
2.1 Méthodes de descente
A partir d’une solution trouvée par heuristique par exemple, on peut très faci-
lement implémenter des méthodes de descente. Ces méthodes s’articulent toutes
60
2.1 Méthodes de descente
autour d’un principe simple. Partir d’une solution existante, chercher une solu-
tion dans le voisinage et accepter cette solution si elle améliore la solution cou-
rante.
L’algorithme 2.1 présente le squelette d’une méthode de descente simple. A
partir d’une solution initiale x, on choisit une solution x 0 dans le voisinage N (x)
de x. Si cette solution est meilleure que x, ( f (x 0 ) < f (x)) alors on accepte cette
solution comme nouvelle solution x et on recommence le processus jusqu’ à ce
qu’il n’y ait plus aucune solution améliorante dans le voisinage de x.
Algorithme 2.1 Simple descent

1: initialise : find an initial solution x
2: repeat
3: neighbourhood search : find a solution x 0 ∈ N (x)
4: if f (x 0 ) < f (x) then
5: x0 ← x
6: end if
7: until f (y) ≥ f (x), ∀y ∈ N (x)
Une version plus “agressive” de la méthode de descente est la méthode de

plus grande descente. Au lieu de choisir une solution x 0 dans le voisinage de x,
on choisit toujours la meilleure solution x 0 du voisinage de x. L’algorithme 2.2
donne une description de cette méthode.
Algorithme 2.2 Deepest descent

2: repeat
3: neighbourhood search : find a solution x 0 ∈ N (x) / f (x 0 ) ≤ f (x 00 ),
∀x 00 ∈ N (x)
4: if f (x 0 ) < f (x) then
5: x0 ← x
6: end if
7: until f (x 0 ) ≥ f (x), ∀x 0 ∈ N (x)
Ces deux méthodes sont évidemment sujettes à de nombreuses critiques. Elles

se basent toutes les deux sur une amélioration progressive de la solution et donc
resteront bloquées dans un minimum local dès qu’elles en rencontreront un. Il
existe de manière évidente une absence de diversification. L’équilibre souhaité
entre intensification et diversification n’existe donc plus et l’utilisateur de ces
deux méthodes doit en être conscient.
Un moyen très simple de diversifier la recherche peut consister à re-exécuter
un des algorithmes en prenant un autre point de départ. Comme l’exécution de
ces méthodes est souvent très rapide, on peut alors inclure cette répétition au
61
sein d’une boucle générale. On obtient alors un algorithme de type “Multi-start

descent” décrit par l’algorithme 2.3.
Algorithme 2.3 Multistart descent

1: initialise : find an initial solution x, k ← 1, f (B) ← +∞
2: repeat
3: Starting point : choose an initial solution x 0 at random
4: x ← result of Simple Descent or Deepest Descent
5: if f (x) < f (B) then
6: B←x
7: end if
8: k ← k+1
9: until stopping criterion satisfied
Dans la section 2.8, nous verrons que cette dernière méthode peut dans cer-
tains cas être particulièrement efficace. De manière évidente, la diversification est
totalement absente des algorithmes 2.1 et 2.2. En ne conservant que l’aspect in-
tensification, la convergence est souvent trop rapide (plus encore avec la version
Deepest descent) et on se trouve très rapidement bloqué dans un optimum local.
Les résultats communément admis indiquent que ces techniques conduisent en
général à des solutions en moyenne à 20% de l’optimum. Dans le cas de l’algo-
rithme 2.3, la diversification est simplement insérée par le choix aléatoire d’une
solution de départ. Cette simple technique peut s’avérer payante dans certains
cas (voir section 2.8).
2.2 Recuit simulé
La méthode du recuit simulé a été introduite en 1983 par Kirkpatrick et al.

[64]. Cette méthode originale est basée sur les travaux bien antérieurs de Metro-
polis et al. [81]. Cette méthode que l’on pourrait considérer comme la première
métaheuristique “grand public” a reçu l’attention de nombreux travaux et prin-
cipalement de nombreuses applications.
Le principe de fonctionnement s’inspire d’un processus d’amélioration de la
qualité d’un métal solide par recherche d’un état d’énergie minimum correspon-
dant à une structure stable de ce métal. L’état optimal correspondrait à une struc-
ture moléculaire régulière parfaite. En partant d’une température élevée où le
métal serait liquide, on refroidit le métal progressivement en tentant de trouver
le meilleur équilibre thermodynamique. Chaque niveau de température est main-
tenu jusqu’à obtention d’un équilibre. Dans ces phases de température constante,
on peut passer par des états intermédiaires du métal non satisfaisants, mais con-
duisant à la longue à des états meilleurs.
L’analogie avec une méthode d’optimisation est trouvée en associant une so-
62
2.2 Recuit simulé
lution à un état du métal, son équilibre thermodynamique est la valeur de la fonc-

tion objectif de cette solution. Passer d’un état du métal à un autre correspond à
passer d’une solution à une solution voisine.
Pour passer à une solution voisine, il faut respecter l’une des deux conditions :
– soit le mouvement améliore la qualité de la solution précédente, i.e. en mi-
nimisation la variation de coût est négative (∆C < 0),
– soit le mouvement détériore la qualité de la solution précédente et la proba-
bilité p d’accepter un tel mouvement est inférieure à une valeur dépendant
de la température courante t (p < e −∆C/t ).
Le schéma de refroidissement de la température est une des parties les plus
difficiles à régler dans ce cas. Ces schémas sont cruciaux pour l’obtention d’une
implémentation efficace. Sans être exhaustif, on rencontre habituellement trois
grandes classes de schémas : la réduction par paliers, la réduction continue, Lundy
et Mees [78] (où la température est réduite à chaque itération) et la réduction non-
monotone, Connolly [15] (où des augmentations de température sont possibles).
Pour d’autres schémas de refroidissement de température et une large discussion,
on peut se reporter à plusieurs articles, dont les travaux de Aarts et Korst [1], de
Collins et al. [14] et de Van Laarhoven et Aarts [125] par exemple. La convergence
du recuit simulé vers un optimum global, en choisissant le schéma adapté, a pu
être prouvée (voir [1, 55, 125]).
La popularité du recuit simulé a été incontestable pendant des années. D’abord
cette méthode est facile à implémenter et elle a permis de résoudre de nom-
breux problèmes N P-difficiles [9, 126]. Parmi les bibliographies et articles de
synthèse intéressants, on peut citer Koulamas et al. [65] et Collins et al. [14]. Un
excellent tutorial récent pourra guider les chercheurs dans leurs premiers pas
avec cette méthode [90]. En attendant, l’algorithme 2.4 présente les principales
caractéristiques d’un recuit simulé.
Algorithme 2.4 Basic simulated annealing

1: initialise : find an initial solution x, fix an annealing schedule T , set initial
temperature t
2: repeat
3: neighbourhood search : find a solution x 0 ∈ N (x)
4: determine ∆C = f (x 0 ) − f (x)
5: draw p ∼ U (0, 1)
6: if ∆C < 0 or e−∆C/t > p then
7: x0 ← x
8: end if
9: update temperature t according to T
Dans cet algorithme, l’équilibre entre intensification et diversification est res-

pecté. L’intensification peut se faire par l’intermédiaire des étapes des lignes 3, 6
63
et 9 de l’algorithme 2.4. A la ligne 3, on effectue une recherche des voisins et ceci

peut être fait par une recherche locale. De plus, à la ligne 6, on accepte tous les
mouvements strictement améliorants comme dans une méthode de descente. Le
schéma de refroidissement de la température réduit lui-aussi la probabilité d’ac-
cepter des mouvements non-améliorants et intensifie la recherche. Par ailleurs, la
diversification est obtenue en autorisant des mouvements non-améliorants sous
condition de probabilité (ligne 6) ou en choisissant un schéma T (ligne 9) permet-
tant l’augmentation de température, donc en augmentant la probabilité d’accep-
tation de solutions plus diverses.
2.3 Recherche tabou
Dans un article présenté par Glover [37] en 1986, on voit apparaı̂tre pour la
première fois, à la fois le terme tabu search et métaheuristique. A la même époque,
Hansen [56] présente une méthode similaire, mais dont le nom n’a pas marqué
autant que tabou. En fait, les prémices de la méthode ont été présentés initiale-
ment à la fin des années 1970 par Glover [36]. Pourtant ce sont les deux articles
de référence de Glover [38, 39] qui vont contribuer de manière importante à la
popularité de cette méthode. Pour certains chercheurs, elle apparaı̂t même plus
satisfaisante sur le plan scientifique que le recuit simulé, car la partie “aléatoire”
de la méthode a disparu.
Contrairement au recuit simulé qui ne génère qu’une seule solution x 0 “aléa-
toirement” dans le voisinage N (x) de la solution courante x, la méthode tabou,
dans sa forme la plus simple, examine le voisinage N (x) de la solution courante x.
La nouvelle solution x 0 est la meilleure solution de ce voisinage (dont l’évaluation
est parfois moins bonne que x elle-même). Pour éviter de cycler, une liste tabou
(qui a donné le nom à la méthode) est tenue à jour et interdit de revenir à des
solutions déjà explorées. Dans une version plus avancée de la méthode tabou,
on peut voir dans cette recherche une modification temporaire de la structure de
voisinage de la solution x permettant de quitter des optima locaux. Le voisinage
N ∗ (x) intégrant ces modifications de structure est régit par l’utilisation de struc-
tures de mémoire spécifiques. Il s’agit de mémoire à court terme ou de mémoire
à long terme.
La mémoire à court terme correspond à la mise en place d’une liste tabou.
La liste contient les quelques dernières solutions qui ont été récemment visitées.
Le nouveau voisinage N ∗ (x) exclut donc toutes les solutions de la liste tabou.
Lorsque la structure de donnée correspondant aux solutions est trop complexe
ou occupe une grande place mémoire, il est courant de ne garder dans la liste
tabou que des informations soit sur les caractéristiques des solutions, soit sur les
mouvements. Ce type de mémoire à court terme est aussi appelé recency-based
memory. En conservant des caractéristiques des solutions ou des mouvements, il
est possible alors qu’une solution de bien meilleure qualité ait un statut tabou.
64
2.3 Recherche tabou
Accepter tout de même cette solution revient à outrepasser son statut tabou, c’est
l’application du critère d’aspiration. Si le voisinage d’une solution est tr ès grand,
évaluer toutes les solutions de ce voisinage peut-être impossible. Il convient alors
de mettre en place des stratégies permettant sa réduction. Les solutions les plus
courantes proposent des listes de solutions candidates qui pourraient conduire à
des solutions de bonne qualité (candidate list strategy).
La mémoire à long terme permet d’une part d’éviter de rester dans une seule
région de l’espace de recherche et d’autre part d’étendre la recherche vers des
zones plus intéressantes. Par exemple, la frequency-based memory ou mémoire à
base de fréquence attribue des pénalités à des caractéristiques des solutions plu-
sieurs fois visitées au cours de la recherche. Cette technique simple permet de
diversifier la recherche facilement. Par ailleurs, les mouvements ayant conduit a
des bonnes solutions peuvent être aussi encouragés. On peut par exemple gar-
der en mémoire une liste de solutions élites que l’on utilisera comme nouveau
point de départ quand la recherche deviendra improductive pendant plusieurs
itérations consécutives.
Il existe aujourd’hui un très grand nombre de références sur cette méthode.
Un des résultats les plus attendus et pourtant ayant un impact limité en pra-
tique est la preuve de convergence de la méthode proposée par Glover et Hanafi
[44]. Parmi les articles introductifs intéressants, on peut citer celui de De Werra et
Hertz [20] et de Glover [40]. Une autre référence particulièrement adaptée se doit
de figurer ici, Hertz et al. [59]. Plus récemment le livre très complet de Glover et
Laguna [45] donne le détail de toutes les techniques citées précédemment et bien
d’autres.
Algorithme 2.5 Basic tabu search

2: repeat
3: neighbourhood search : find a solution x 0 ∈ N ∗ (x)
4: update memory : tabu list, frequency-based memory, aspiration level, . . .
5: move x ← x 0
L’algorithme 2.5 présente une version très simplifiée d’un algorithme de re-
cherche tabou. L’avantage de cette présentation, c’est qu’elle n’exclut aucune des
extensions possibles mais évidemment avec un peu moins de clarté. L’alternance
entre intensification et diversification est bien respectée. L’intensification réside
dans la technique de base qui est l’acceptation d’une solution améliorante et bien
sûr par le critère d’aspiration. Dans la mémoire à long terme, la conservation de
solutions élites est aussi un facteur d’intensification. Pour l’aspect diversification,
le rôle le plus important est joué par la liste tabou elle-même, mais aussi par cer-
tains points de la mémoire à long terme (la frequency-based memory par exemple).
65
2.4 Recherche à voisinages variables
La recherche à voisinages variables est une méthode récente et pourtant très

simple, basée sur la performance des méthodes de descente. Introduite par Mla-
denović et Hansen [85], la méthode propose simplement d’utiliser plusieurs voi-
sinages successifs quand on se trouve bloqué dans un minimum local.
Avant tout, il est nécessaire de définir un ensemble de k max voisinages, dénotés
par Nk=1...k max (et de préférence tels que Nk ⊂ Nk+1 ). On choisit une solution de
départ x par heuristique. Ensuite, à partir d’une solution initiale x 0 choisie dans
le premier voisinage N (x) de x, on applique une méthode de descente (ou une
autre méthode de recherche locale) jusqu’à arriver dans un minimum local (ou
que la recherche locale s’arrête). Si la solution trouvée x 00 est meilleure que x alors
on recentre la recherche en repartant du premier voisinage, sinon on passe au
voisinage suivant (qui a priori est plus grand). La recherche s’arrête quand tous
les voisinages ne sont plus capables d’améliorer la solution.
Le point crucial dans une VNS, c’est bien évidement la constitution des voisi-
nages de plus en plus grands et inclus les uns dans les autres. Une bonne structure
de voisinage –ceci étant vrai pour toute recherche locale– conduit généralement à
de bons résultats ou au moins à une recherche efficace.
Cette méthode récente intéresse de plus en plus de chercheurs et est promise
à un bel avenir. Le nombre de papiers utilisant cette méthode est en très forte
augmentation dans les congrès. La méthode est décrite en détail par Hansen et
Mladenović [57] et un squelette général de cette méthode est proposé dans l’algo-
rithme 2.6.
Algorithme 2.6 Basic variable neighbourhood search

1: initialise : find an initial solution x, k ← 1
2: repeat
3: shake : generate a point x 0 at random from the neighbourhood N k (x)
4: local search : apply a local search procedure starting from the solution x 0 to
find a solution x 00
5: if x 00 is better than x then
6: x ← x 00 and k ← 1 (centre the search around x 00 and search again with a
small neighbourhood)
7: else
8: k ← k + 1 (enlarge the neighbourhood)
9: end if
10: until k = k max
Dans l’algorithme 2.6, la diversification est gérée par deux techniques. La

première consiste à choisir dans le voisinage courant une solution aléatoirement
(étape shake) et la seconde c’est le changement de voisinage lui-même qui agrandit
66
2.5 GRASP
l’espace de recherche autorisé et donc agrandit le voisinage exploré. L’intensifica-

tion de la recherche est effectuée par l’appel à une procédure de recherche locale.
Le plus simple est d’implémenter une méthode de descente (intensification uni-
quement), mais parfois on peut souhaiter appliquer à cet endroit une recherche
plus évoluée comme une recherche tabou qui contient elle aussi des techniques
de diversification (voir par exemple Gagne et al. [33]).
2.5 GRASP
Introduite en 1989 par Feo et Resende [26] et présentée dans sa forme plus
définitive en 1995 [27], la méthode Greedy Randomized Adaptative Search (GRASP)
combine une heuristique gloutonne et une recherche aléatoire. A chaque itération,
on construit une solution comme dans une heuristique gloutonne (en se servant
d’une liste d’attributs comme liste de priorité). Cette solution est améliorée par
l’intermédiaire d’une méthode de descente. En se basant sur la qualité générale
de la solution ainsi obtenue, on met à jour l’ordre de la liste des attributs et
le processus est itéré jusqu’à satisfaction d’un critère d’arrêt. Un des avantages
de cette méthode est la simplicité avec laquelle on peut comprendre le proces-
sus d’optimisation. La mise en œuvre elle aussi n’est pas trop compliqu ée. Des
implémentations réussies et un tutorial récent présentant cette méthode sont pré-
sentées par Festa [28].
Algorithme 2.7 Basic GRASP

1: construct an ordered list of solution attributes S
2: repeat
3: repeat
4: take s ∈ S in a greedy way
5: x = x∪s
6: until solution x complete
7: local search : find a locally optimal solution x 0 starting from x
8: update the benefits of the solution attributes in S
Dans l’algorithme 2.7, on retrouve les deux phases indispensables à la méthode

au sein de la boucle repeat (lignes 2 à 9). La construction d’une solution par l’al-
gorithme glouton se fait par la boucle interne repeat (lignes 3 à 6). L’étape de mise
à jour de la liste des attributs n’est pas anodine pour la méthode et se trouve être
un des points clés pour la réussite de cette dernière. L’intensification est obtenue
par l’application de la recherche locale au sein de la boucle principale. Mais la
mise à jour de la liste des attributs est aussi un facteur important permettant de
contrôler la convergence de la méthode et donc de son intensification. Mettre en
évidence l’aspect diversification de cette méthode n’est pas simple, pourtant c’est
le rôle de l’algorithme glouton. De par sa nature, cet algorithme glouton peut
67
commencer à construire une bonne solution, mais en effectuant un choix aveugle

à chaque étape de la construction d’une solution, il peut s’éloigner rapidement
d’une bonne solution très proche. C’est donc un facteur de diversification dans la
méthode GRASP.
2.6 Iterated local search
La méthode iterated local search est une variante très simple des méthodes de
descente qui pallient au problème de l’arrêt de ces dernières dans des optima
locaux. Donner la paternité de cette méthode à un ou plusieurs auteurs serait
abusif. On peut tout de même citer dans l’ordre chronologique quelques unes
des références présentant des méthodes très proches : Baxter [5], Baum [4], Mar-
tin et al. [80]. Par contre, le nom de la méthode pourrait revenir sans conteste à
Lourenço et al. [76] (voir aussi Lourenço et al. [77]).
Dans cette méthode, on génère une solution initiale qui servira de point de
départ. Ensuite, on va répéter deux phases : une phase de perturbation aléatoire
dans laquelle la solution courante va être perturbée (parfois en tenant compte
d’un historique maintenu à jour) et une seconde phase de recherche locale (ou
tout simplement de méthode de descente) qui va améliorer cette solution jusqu’à
buter sur un optimum local. Dans cette boucle, on est aussi à même d’accepter
ou non la nouvelle solution selon que l’on souhaite donner un caractère plus ou
moins agressif à la méthode.
Pour une présentation plus pragmatique de la méthode nous orientons le lec-
teur vers den Besten et al. [19]. L’implémentation détaillée de la méthode pour
résoudre un problème d’ordonnancement à une machine y est présentée.
Algorithme 2.8 Basic iterated local search

2: repeat
3: random perturbation : find a solution x 0 “close” to x
4: local search : starting from x 0 , find a solution x 00 locally optimal
5: if x 00 is accepted (e.g. better than x) then
6: x ← x 00
7: end if
L’algorithme 2.8 présente un squelette général de la méthode iterated local

search. Sans ambiguı̈té, la diversification est obtenue par la perturbation aléatoire.
Même si cette dernière peut dans certaines implémentations tenir compte d’un
historique des perturbations, il est important de conserver ici le caract ère aléatoire.
Encore une fois, c’est une recherche locale qui intensifie la recherche. La phase
d’acceptation de la solution au sein de la boucle générale est importante. On peut
68
2.7 Guided local search
par exemple accepter une solution si elle améliore strictement la solution cou-
rante. Mais on peut aussi accepter seulement si elle améliore la meilleure solution
trouvée, ou pourquoi pas, si elle ne dégrade pas la meilleure solution trouvée de
plus d’un certain pourcentage. Comme on le voit, on peut adapter cette méthode
facilement et ajouter des facteurs d’intensification ou de diversification à cet en-
droit.
2.7 Guided local search
Présentée pour la première fois dans un rapport de recherche de Voudouris

et Tsang [128], la guided local search (ou recherche locale guidée) est une variante
assez élaborée d’une méthode de descente classique.
La méthode de base est simple. Elle consiste à modifier la fonction à optimiser
en ajoutant des pénalités. La recherche locale est appliquée alors sur cette fonction
modifiée. La solution trouvée (qui se trouve être un optimum local) sert à calculer
les nouvelles pénalités. Pour cela, on calcule l’utilité de chacun des attributs de la
solution et on augmente les pénalités associées aux attributs de valeur maximale.
Ces étapes successives sont répétées jusqu’à ce qu’un critère d’arrêt soit validé.
L’article initial de Voudouris et Tsang [129], soumis en octobre 1995 n’a été pu-
blié qu’en 1999 et présente de nombreuses améliorations de l’algorithme 2.9. Pour
une présentation très approfondie, la thèse de Voudouris [127] est le document le
plus complet.
Algorithme 2.9 Basic guided local search

1: initialise : find an initial solution x, set penalties to 0
2: repeat
3: cost function : augment the original cost function of x with penalties
4: local search : apply a local search procedure to find a solution x 0 locally op-
timal (based on the augmented cost function)
5: update penalties : by computing utility expressions
6: x ← x0
Dans l’algorithme 2.9, nous avons volontairement simplifié la présentation

pour ne pas surcharger la lecture et la compréhension. L’intensification de la re-
cherche est fait directement par l’appel à la recherche locale au sein de la boucle
principale. Les expressions d’utilités des différents attributs d’une solution servent,
pour celles qui sont à leur maximum, à pénaliser la fonction de coût. Ce qui veut
dire que l’on va forcer la recherche à s’effectuer dans d’autres directions que celles
qui semblaient prometteuses lors des précédentes recherches. Cette mise à jour
des fonctions de pénalité est donc un facteur de diversification de la recherche.
69
2.8 Applications
Applications de la littérature
Comme il a été mentionné précédemment, il existe assez peu de références qui
ne présentent que des méthodes de descente. Pourtant, certains ouvrages plus
complet ou dépassant ce cadre sont intéressants (voir par exemple Blum et Roli
[8] ou Johnson et al. [63]). Il existe énormément d’applications utilisant le recuit
simulé. On peut citer des articles importants de Van Laarhoven et Aarts [125] ou
de Vidal [126] et la bibliographie commentée de Collins et al. [14] qui référence un
très grand nombre d’applications. Pour la recherche tabou, on ne compte plus les
articles se référant à cette méthode. On peut citer un numéro spécial de Glover
et al. [47] présentant plusieurs articles sur le sujet. La méthode GRASP est de
plus en plus populaire. Il existe un site web http://www.graspheuristic.org
qui est maintenu à jour et référence plusieurs centaines d’articles. Régulièrement
une bibliographie commentée de GRASP est publiée (voir Festa et Resende [29]).
La méthode Iterated local search, un peu ignorée au début, sans doute pour son
apparente simplicité, commence à prendre place au sein de toutes ces méthodes
(voir Paquete et Stützle [89]). Du côté du Guided local search, c’est le même constat.
Les applications se multiplient et la méthode prouve son efficacité dans plusieurs
domaines (voir Beullens et al. [7] ou Voudouris et Tsang [129]).
Ordonnancement sur machines parallèles

De notre côté, nous avons abordé plusieurs de ces méthodes pour un problème
d’ordonnancement sur machines parallèles. Un ensemble de n jobs (j = 1 . . . n)
doit être séquencé sur m machines parallèles. Chaque job doit être exécuté sur une
des machines et ne peut démarrer avant sa date de début au plus tôt, r j (release
date). Le temps d’exécution de ce job est de p j unités de temps (processing time).
Une date échue, d j (due date) et un poids, w j (weight) sont donnés pour chaque
job. Si le job se termine après sa date échue, on dit qu’il est en retard, sinon le
job est terminé à temps. Le critère retenu est la minimisation de la somme des
poids des jobs en retard min ∑ w j Uj . Ce problème est noté Pm|r j | ∑ w j Uj dans
la littérature [51]. Même réduit à deux machines, ce problème est N P-difficile
[35]. Ce problème et les résultats présentés plus loin sont extraits des différentes
présentations du problème (voir par exemple [113–118]).
Le tableau 1 est directement extrait de [115]. Nous comparons ici les méthodes
en utilisant toujours le même voisinage. Dans ce cas-ci, il s’agit d’échanger un job
actuellement en retard par un ou plusieurs jobs en avance (voir Sevaux et Tho-
min [115]). Sans entrer dans des détails, les lignes notées EMP, BWS, BNO, BRS
représentent différents points de départ des heuristiques (sans objet pour la mé-
thode multi-start).
Sans tirer des conclusions générales, on constate dans le tableau 1 que le point
de départ a peu d’impact sur la qualité d’une solution obtenue par méthode de
70
TAB . 1 – Résultats extraits de Sevaux et Thomin [115].
Initial Optimum First Gap to Avg. CPU Max. CPU
Heuristics hits (%) pos. (%) opt. (%) time (s) time (s)
Deepest descent heuristics
EMP 45 33 21.5 0.02 0.12
BWS 49 35 19.6 < .01 < .01
BNO 44 31 20.3 < .01 < .01
BRS 44 31 20.3 < .01 < .01
Multistart deepest descent heuristic
MD 78 63 3.9 4.58 41.1
Simulated annealing
EMP 70 55 6.72 9.10 60.3
BWS 70 54 6.59 7.68 53.2
BNO 71 56 6.03 9.01 64.9
BRS 70 55 5.98 8.87 63.3
Tabu search
EMP 89 79 1.94 10.63 101.9
BWS 88 76 2.00 9.42 112.0
BNO 88 80 1.89 10.17 93.7
BRS 88 79 1.95 10.16 107.5
descente et que les solutions se trouvent en moyenne à 20% de l’optimum. En uti-

lisant le même voisinage, et pour une configuration particulière des paramètres
de recuit simulé et recherche tabou, les écarts à l’optimum sont respectivement
de 6% et 2% environ. Par contre, ce qui est très surprenant, c’est que la méthode
multi-start obtient de très bons résultats avec un écart de 4%. Faut-il pour au-
tant renoncer à implémenter un recuit simulé ? Certainement pas. Comme il a été
signalé, le réglage des paramètres est fixe et en faisant varier ces derniers on de-
vrait pouvoir atteindre des résultats meilleurs. Ce qui est important à retirer de
cette expérience, c’est que même si l’implémentation des méthodes de descente
donne des résultats médiocres, en passant à une simple méthode multi-start on
obtient des solutions très satisfaisantes avec des temps de calcul raisonnables et
surtout un temps d’implémentation minimum (le plus dur étant déjà fait dans
l’écriture du voisinage lui-même). Donc, passer ensuite à une méthode plus com-
plexe comme le recuit simulé ou la recherche tabou est un pas supplémentaire à
franchir pour peu qu’on en ait le temps.
3 Métaheuristiques à base de population

Les méthodes de recherche à population, comme leur nom l’indique, travaillent
sur une population de solutions et non pas sur une solution unique. On peut
71
trouver d’autres noms génériques pour ces méthodes, le plus en vogue étant
sans doute evolutionary algorithms (voir Michalewicz [82]). Le principe général de
toutes ces méthodes consiste à combiner des solutions entre elles pour en former
de nouvelles en essayant d’hériter des “bonnes” caractéristiques des solutions
parents. Un tel processus est répété jusqu’à ce qu’un critère d’arrêt soit satisfait
(nombre de générations maximum, nombre de générations sans améliorations,
temps maximum, borne atteinte, etc). Parmi ces algorithmes à population, on re-
trouve deux grandes classes qui sont les algorithmes génétiques et les colonies
de fourmis. Les algorithmes génétiques ont beaucoup fait parler d’eux et de-
puis longtemps. Les colonies de fourmis sont des techniques plus récentes, mais
l’auteur a su faire parler de la méthode et intéresser non- seulement les cher-
cheurs, mais aussi les grandes instances décisionnaires comme la commission eu-
ropéenne par exemple. Il existe bien d’autres techniques que nous aurions pu
décrire ici comme le Genetic programming de Koza [67] par exemple. Nous aver-
tissons aussi ici le lecteur, que les prises de position sur différents points sont
celles de l’auteur de cette synthèse et n’engage que lui. Elles sont pour la plupart
discutables...
3.1 Algorithmes génétiques
Proposé dans les années 1975 par Holland [61], les algorithmes génétiques
doivent leur popularité à Goldberg [49]. Avant la parution de son livre qui est
une des références les plus citées dans le domaine de l’informatique, on a pu
voir un certain nombre d’autres présentations, citons Goldberg [48], Holland [62],
Schwefel [97]. Le sujet connaı̂t une très grande popularité. Il existe aujourd’hui
plusieurs milliers de références sur le sujet et le nombre de conférences dédiées
au domaine (que ce soit sur les techniques elles-mêmes ou sur les applications)
ne fait qu’augmenter.
De manière générale, les algorithmes génétiques utilisent un même principe.
Une population d’individus (correspondants à des solutions) évoluent en même
temps comme dans l’évolution naturelle en biologie. Pour chacun des individus,
on mesure sa faculté d’adaptation au milieu extérieur par le fitness. Les algo-
rithmes génétiques s’appuient alors sur trois fonctionnalités :
la sélection qui permet de favoriser les individus qui ont un meilleur fitness (pour
nous le fitness sera la plus souvent la valeur de la fonction objectif de la so-
lution associée à l’individu).
le croisement qui combine deux solutions parents pour former un ou deux en-
fants (offspring) en essayant de conserver les “bonnes” caractéristiques des
solutions parents.
la mutation qui permet d’ajouter de la diversité à la population en mutant cer-
taines caractéristiques (gènes) d’une solution.
72
3.1 Algorithmes génétiques
La représentation des solutions (le codage) est un point critique de la réussite

d’un algorithme génétique. Il faut bien sûr qu’il s’adapte le mieux possible au
problème et à l’évaluation d’une solution. Le codage phénotypique ou codage
direct correspond en général à une représentation de la solution très proche de
la réalité. L’évaluation d’une solution représentée ainsi est en général immédiate.
On peut aussi utiliser un codage indirect (codage génotypique) qui est souvent
plus éloigné de la réalité et qui nécessite un algorithme de décodage pour recons-
tituer une solution valide. Pour ce dernier codage, il existe aujourd’hui beaucoup
de travaux et notamment des opérateurs de croisement et de mutation en quan-
tité.
En effectuant une recherche sur internet, on trouve plusieurs milliers de pages
web répondant à la requête genetic algorithms. Donner une bonne référence est in-
dispensable. Nous renvoyons donc le lecteur vers deux livres importants, Haupt
et Haupt [58], Mitchel [84] et vers un bon article introductif qui a le mérite d’être
en français, Fleurent et Ferland [30] et qui explique avec suffisamment de détails
une implémentation parmi d’autres. Comme pour d’autres algorithmes, la preuve
de convergence est apportée par Cerf [12]. Même si l’intérêt pratique de cette
preuve est limité, il est toujours plus satisfaisant de savoir que la méthode converge
sous certaines conditions.
Nous présentons deux versions possibles d’algorithme génétique. L’algorithme
3.1 propose une version qualifiée en Anglais par population replacement (ou algo-
rithme générationnel) tandis que l’algorithme 3.2 est une version incrémentale.
La différence entre les deux réside dans la gestion de la population.
Dans le premier cas (algorithme 3.1), on créé une seconde population en pa-
rallèle à la première, puis on insère des individus obtenus par croisement et muta-
tion des individus de la population précédente. Dès qu’une nouvelle population
est formée, elle remplace la population précédente et le processus recommence.
Un des points importants à prendre en compte est la perte possible du meilleur
individu ou même de la solution optimale lors du remplacement de la popu-
lation. Il faudra donc prévoir un mécanisme permettant d’éviter cela. Dans cet
algorithme, la sélection est faite ici en effectuant un tirage aléatoire biaisé par
la valeur du fitness des individus (par exemple par la méthode de ranking de
Reeves [95]). Comme dans les systèmes d’évolution naturels, le croisement et la
mutation se font sous conditions de probabilité.
Dans la seconde version de l’algorithme génétique (algorithme 3.2), on créé
un nouvel individu par croisement et mutation et il remplace un individu exis-
tant dans la population courante. Ma préférence va à la seconde méthode qui
permet (potentiellement) d’exploiter immédiatement tout nouvel individu créé à
l’itération précédente. Dans cet algorithme, nous choisissons la méthode de tour-
noi binaire (ou roulette wheel) pour la sélection. Le croisement est fait à chaque
itération et la mutation sous conditions (par exemple avec une faible probabilit é
mais en respectant un critère d’aspiration). Pour le choix de l’individu à rem-
73
Algorithme 3.1 A simple genetic algorithm (population replacement)

1: initialise : generate an initial population P of solutions with size |P| = n
2: repeat
3: P0 ← ∅
4: repeat
5: selection : choose 2 solutions x and x 0 from P with probability proportio-
nal to their fitness
6: crossover : combine parent solutions x and x 0 to form child solutions y and
y0 with high probability
7: mutate y and y0 with small probability
8: add y and y0 to P0
9: until |P0 | = n
10: P ← P0
placer, on peut comme dans Reeves [95] choisir un individu parmi ceux dont le
fitness dépasse la valeur médiane (plus mauvaise moitié des individus) ou alors
utiliser la technique de tournoi binaire (inversée). On peut bien sûr modifier ce
squelette facilement avec toutes les adaptations proposées dans la littérature.
Algorithme 3.2 A simple genetic algorithm (incremental replacement)

2: repeat
3: selection : choose 2 solutions x and x 0 with the roulette wheel technique
4: crossover : combine parent solutions x and x 0 to form a child solution y
5: mutate y under conditions
6: choose an individual y 0 to be replaced in the population
7: replace y0 by y in the population
Pour l’un et l’autre des algorithmes, l’intensification est amenée à la fois par
le processus de sélection et par l’opérateur de croisement. Comme la sélection
est faite soit en biaisant un processus aléatoire, soit par tournoi binaire (favo-
risant dans les deux cas les meilleurs individus), on va forcer la population à
converger. C’est donc un facteur d’intensification. Les opérateurs de croisement
sont dessinés pour faire profiter aux enfants des “bonnes” caractéristiques des
parents. Même si ce n’est pas toujours évident de construire de tels opérateurs,
ces derniers devraient constituer un facteur d’intensification. Nous verrons dans
le cas des algorithmes mémétiques (section 4.1), que le croisement peut être aussi
vu comme un facteur de diversification dans certains cas. Mais dans les algo-
rithmes génétiques initiaux, le facteur de diversification principal est la mutation
qui modifie directement les caractéristiques d’une solution permettant d’explorer
74
3.2 Algorithmes de colonies de fourmis
de nouveaux espaces de solutions.
3.2 Algorithmes de colonies de fourmis
Les algorithmes à base de colonies de fourmis ont été introduits par Dorigo
et al. [24] et dans la thèse de [21]. Une des applications principales de la méthode
originale était le problème du voyageur de commerce et depuis elle a considéra-
blement évolué.
Cette nouvelle métaheuristique imite le comportement de fourmis cherchant
de la nourriture. A chaque fois qu’une fourmi se déplace, elle laisse sur la trace de
son passage une odeur (la phéromone). Comme la fourmi est rarement une explo-
ratrice isolée, avec plusieurs de ses congénères, elle explore une région en quête
de nourriture. Face à un obstacle, le groupe des fourmis explore les deux côtés
de l’obstacle et se retrouvent, puis elles reviennent au nid avec de la nourriture.
Les autres fourmis qui veulent obtenir de la nourriture elles aussi vont emprun-
ter le même chemin. Si celui-ci se sépare face à l’obstacle, les fourmis vont alors
emprunter préférablement le chemin sur lequel la phéromone sera la plus forte.
Mais la phéromone étant une odeur elle s’évapore. Si peu de fourmis empruntent
une trace, il est possible que ce chemin ne soit plus valable au bout d’un mo-
ment, il en est de même si des fourmis exploratrices empruntent un chemin plus
long (pour le contournement de l’obstacle par exemple). Par contre, si le chemin
est fortement emprunté, chaque nouvelle fourmi qui passe redépose un peu de
phéromone et renforce ainsi la trace, donnant alors à ce chemin une plus grande
probabilité d’être emprunté.
On peut alors comprendre l’application directe du problème du voyageur
de commerce. Chaque fourmi constitue un tour en empruntant des arêtes du
graphes, puis au bout d’un moment les arêtes qui conduisent à une bonne so-
lution deviendront les plus empruntées jusqu’à obtention d’une très bonne so-
lution. Même si le principe est simple, on peut voir apparaı̂tre des difficultés
comme, la quantité de phéromone à déposer, le coefficient d’évaporation, la dé-
finition du choix biaisé, etc. La métaheuristique telle qu’elle est décrite schémati-
quement ci- dessous est détaillée par Dorigo et al. [23] et Dorigo et Di Caro [22].
Bien que cette méthode soit plutôt récente, elle a suscité de nombreux articles en
“tournées de véhicules”, peut-être par son exotisme ou parce qu’il est plus facile
de publier avec une nouvelle méthode moins connue.
Si on regarde en détail l’algorithme 3.3, on pourrait le comparer à l’algorithme
GRASP (voir section 2.5) dans lequel on ferait entrer en jeu plusieurs solutions
simultanément. Chaque fourmi est associée à un algorithme glouton biaisé où les
phéromones sont associées aux attributs de GRASP. Pour parfaire la comparaison,
il manque la recherche locale de GRASP qui n’est pas présente ici, mais que l’on
commence à voir apparaı̂tre dans des exposés en conférence.
75
Algorithme 3.3 Basic ant colony optimisation

1: initialise : create an initial population of ants
2: repeat
3: for each ant do
4: construct a solution based on the construction procedure, biased by the
pheromone trails
5: update the pheromone trails based on the quality of the solutions found
6: end for
Les facteurs d’intensification et de diversification sont plus difficiles à cer-

ner. Néanmoins, le choix d’un nouvel attribut se fait aléatoirement (même s’il
est biaisé) c’est donc un facteur de diversification. L’intensification se retrouve
dans les paramètres de dépôt et d’évaporation qui concentrent plus ou moins les
teneurs en phéromone des attributs de la solution.
3.3 Applications
Pour des présentations générales de méthodes à population qui incluent des

applications (principalement en ordonnancement), nous renvoyons le lecteur vers
certains de nos travaux et présentations [99, 102].
Que dire des applications des algorithmes à population. Elles sont de plus en
plus nombreuses, et de plus en plus évoluées. Du côté des techniques de colonies
de fourmis, le sujet est encore un peu jeune pour proposer plusieurs articles de
synthèse sur différents travaux. On peut tout de même citer les articles de Do-
rigo et al. [23] et Dorigo et Di Caro [22]. Pour les algorithmes génétiques, on peut
citer en premier lieu une bibliographie proposée par Alander [2] qui recense des
implémentations. Ce rapport comporte quelques 850 références d’implémentations
des algorithmes génétiques. Le plus impressionnant, c’est que le nombre d’ar-
ticles sur les algorithmes génétiques est d’environ 1000 par an depuis 1995 avec
plus d’une centaine d’articles sur des implémentations chaque année. La différence
entre ces deux chiffres montre qu’il y a encore un potentiel très intéressant pour
de nouvelles applications (voir aussi De Jong [18]). Un site web particulièrement
intéressant présente des articles et implémentations http://www-illigal.ge.
uiuc.edu.
En dehors des ouvrages généraux comme [49, 94] et de tous les autres, on
peut mettre en avant une très bonne introduction aux problèmes résolus par les
algorithmes génétiques présentée par Reeves [96]. Cet article référence des appli-
cations dans tous les domaines de la recherche opérationnelle et en particulier de
76
3.3 Applications
l’optimisation combinatoire.
Ordonnancement à une machine

De notre côté, nous avons mené depuis 1999 un certain nombre de travaux sur les
algorithmes génétiques. Par exemple, dans Sevaux et Dauzère-Pérès [107], nous
proposons une étude assez détaillée d’un algorithme génétique classique pour
lequel nous avons essayé d’appliquer différentes variantes. Le problème concerné
est la minimisation du nombre de jobs en retard sur une machine. Un ensemble
de n jobs (j = 1 . . . n) doit être séquencé sur une machine unique où la préemption
n’est pas autorisée. Chaque job ne peut démarrer avant sa date de début au plus
tôt, r j (release date). Le temps d’exécution de ce job est de p j unités de temps
(processing time). Une date échue, d j (due date) et un poids, w j (weight) sont
donnés pour chaque job. Si le job se termine après sa date échue, on dit qu’il est en
retard, sinon le job est terminé à temps. Le critère retenu est la minimisation de la
somme des poids des jobs en retard min ∑ w j Uj . Ce problème est noté 1|r j | ∑ w j Uj
dans la littérature [51] et N P-difficile au sens fort [75].
Dans cette étude, pour un même codage indirect, nous étudions l’impact de
différents moteurs de décodage de la solution, de plusieurs opérateurs de croi-
sement et de mutation. Ces résultats sont comparés avec des solutions obtenues
par la relaxation Lagrangienne de Dauzère-Pérès et Sevaux [17]. Cette étude est
le fruit de plusieurs expériences (voir [103, 104, 106]).
TAB . 2 – Comparaison pour le moteur “List”, extrait de [105]

Best Indiv. Average Indiv. Cpu Time Av. Nb. of
Mean St-Dev Mean St-Dev (in sec.) Iterations
OX + Swap 4.52 5.53 4.84 5.83 11.30 39309.04
+ BigSwap 4.46 5.46 4.80 5.78 12.79 43156.70
+ Shift 4.53 5.51 4.86 5.81 11.05 37542.16
LOX + Swap 4.53 5.57 4.66 5.68 6.64 23407.71
+ BigSwap 3.83 4.90 4.01 5.05 15.40 55058.94
+ Shift 4.23 5.29 4.39 5.43 10.76 37850.51
PMX+ Swap 4.31 5.41 4.48 5.56 7.37 25536.58
+ BigSwap 3.99 5.09 4.22 5.29 10.68 37171.44
+ Shift 4.00 5.09 4.31 5.36 9.95 34677.41
X1 + Swap 4.81 5.76 4.91 5.85 1.99 14335.95
+ BigSwap 1.01 1.09 1.44 1.58 10.66 78381.29
+ Shift 4.64 5.58 4.73 5.67 2.42 18533.48
La table 2 (resp. 3 et 4) présente les résultats des différentes combinaisons

d’opérateurs de croisement et de mutation obtenus pour un moteur de décodage
d’un individu de type “List” (resp. de type “Sans délais”, et “FIFO”). Les quatre
opérateurs de croisement sont OX, LOX, PMX et X1 et les trois opérateurs de
77
TAB . 3 – Comparaison pour le moteur “ND”, extrait de [105]

OX + Swap 1.58 1.13 2.11 1.55 44.90 26464.21
+ BigSwap 1.48 1.03 2.05 1.49 51.10 29443.37
+ Shift 1.53 1.07 2.08 1.51 46.64 26837.89
LOX + Swap 1.45 1.04 1.72 1.22 37.56 20468.35
+ BigSwap 1.07 0.82 1.46 1.08 57.18 29790.79
+ Shift 1.18 0.87 1.54 1.10 50.12 26661.19
PMX+ Swap 1.21 0.93 1.54 1.16 41.34 22046.54
+ BigSwap 0.89 0.72 1.33 1.00 61.65 31927.69
+ Shift 0.88 0.73 1.45 1.12 58.99 32960.50
X1 + Swap 2.11 1.57 2.30 1.72 11.28 14065.76
+ BigSwap 1.29 0.99 1.58 1.17 26.94 33426.13
+ Shift 1.63 1.15 1.84 1.30 20.28 23572.64
TAB . 4 – Comparaison pour le moteur “Fifo”, extrait de [105]

OX + Swap 1.02 1.01 1.52 1.38 42.61 31572.30
+ BigSwap 0.91 0.91 1.41 1.27 46.81 34969.10
+ Shift 0.98 0.95 1.49 1.32 42.02 31391.71
LOX + Swap 1.12 1.10 1.23 1.17 24.92 18061.56
+ BigSwap 0.72 0.75 0.88 0.86 38.62 28357.40
+ Shift 0.83 0.82 0.98 0.93 33.85 24807.48
PMX+ Swap 0.93 0.97 1.06 1.05 27.91 20047.78
+ BigSwap 0.69 0.74 0.86 0.85 38.65 27862.21
+ Shift 0.69 0.74 0.95 0.95 35.90 27528.53
X1 + Swap 1.76 1.66 1.86 1.73 7.48 13791.81
+ BigSwap 0.74 0.75 0.84 0.83 17.68 32800.92
+ Shift 1.17 1.02 1.25 1.08 13.03 22752.71
78
mutation sont Swap (ou API, Adjacent Pairwise Interchange), BigSwap (ou GPI,
General Pairwise Interchange) ou Shift (insertion d’un gène puis décalage) (pour
plus de détails, voir [105]). Les mesures sont faites par rapport aux solutions op-
timales des problèmes. Les chiffres de ces tableaux sont les moyennes constatées
sur 1000 instances. Les mesures sont effectuées en comparant les écarts (en moyen-
ne et écart-type) du meilleur individu et de l’individu moyen (colonnes “Best In-
div.” et “Average Indiv.”). Les temps CPU et nombres moyens d’itérations sont
donnés à titre indicatif.
De ces tables, on a pu mettre en évidence un avantage certain à un moteur de
décodage, un opérateur de croisement et un opérateur de mutation. En regardant
le critère de performance relative, le moteur de type “List” n’est pas satisfaisant
quelque soient les opérateurs retenus. Le moteur ND de type “Sans délais” donne
de bons résultats quand il est associé à l’opérateur de croisement PMX et pas
avec l’opérateur de mutation “Swap”. Le moteur de type “Fifo” donne de bons
résultats avec les opérateurs de croisement PMX, LOX et X1 et les opérateurs de
mutation “BigSwap” et “Shift”. Pour la suite de l’étude nous avions retenu le
moteur “Fifo”, l’opérateur de croisement PMX et l’opérateur de mutation “Shift”.
4 Métaheuristiques avancées
Pour aller plus loin dans la recherche de solutions, il faut avant tout pouvoir
détecter de nouvelles solutions. Les algorithmes à base de population présentent
un intérêt particulier : le parallélisme intrinsèque. En ajoutant de nouveaux com-
posants à ces algorithmes, on peut alors construire des algorithmes “hybrides”.
L’une des faiblesses d’un algorithme génétique comme la vitesse de convergence
trop lente peut, par exemple, être compensée par l’ajout d’une méthode de re-
cherche locale. C’est le cas des algorithmes mémétiques de Moscato [86] présentés
section 4.1. Il est clair que sans l’aide des recherches locales les méthodes à popu-
lation n’arrivent pas à fournir des solutions très satisfaisantes. De même, sans la
gestion efficace d’une population de solutions, il est difficile pour une recherche
locale de parcourir efficacement l’espace des solutions souvent très vaste. Une
autre des caractéristiques importantes est de pouvoir limiter la taille de la po-
pulation. C’est le cas du “scatter search” de Glover [42] et du nouvel algorithme
GA|PM initialement proposé par Sörensen [121]. Une taxonomie des méthodes
dites hybrides est proposée dans Talbi [124].
4.1 Algorithmes mémétiques
Moscato [86] introduit en 89 pour la première fois les algorithmes mémétiques.

On rencontre aussi le nom d’algorithmes génétiques hybrides (trop vague à notre
goût) ou celui de genetic local search (à ne pas confondre avec guided local search).
79
Quelque soit le nom qu’on lui donne, l’idée principale de cette technique est de
rendre plus agressif un algorithme génétique par l’ajout d’une recherche locale
en plus de la mutation.
Une des observations générales provenant de l’implémentation d’un algo-
rithme génétique basique est souvent la faible vitesse de convergence de l’algo-
rithme. L’idée de Moscato est donc d’ajouter une recherche locale – qui peut être
une méthode de descente ou une recherche locale plus évoluée (recuit simulé ou
recherche tabou par exemple). Cette recherche locale sera appliquée à tout nouvel
individu obtenu au cours de la recherche.
Il est évident que cette simple modification entraı̂ne de profonds changements
dans le comportement de l’algorithme lui-même. Après avoir créé un nouvel in-
dividu à partir de deux parents sélectionnés, on applique une recherche locale et
sous certaines conditions on applique un opérateur de mutation à cet individu.
Les conditions peuvent être une certaine probabilité. Il est aussi possible d’ajouter
un critère d’aspiration ou d’autres techniques plus évoluées à cet endroit.
Pour une introduction aux algorithmes mémétiques, il est souvent plus fa-
cile de se reporter à la page web maintenue à jour par P. Moscato http://www.
densis.fee.unicamp.br/~moscato qui présente les papiers de référence de cette
méthode. Nous suggérons aussi au lecteur de consulter Moscato [87].
Algorithme 4.1 A simple memetic algorithm (incremental replacement)

2: apply a local search procedure on each solution of P
3: repeat
4: selection : choose 2 solutions x and x 0 with the roulette wheel technique
5: crossover : combine parent solutions x and x 0 to form a child solution y
6: local search : apply a local search procedure on y under conditions
7: mutation : apply a mutation operator on y under conditions
8: choose an individual y 0 to be replaced in the population
9: replace y0 by y in the population
Nous proposons une version incrémentale de l’algorithme mémétique (voir

algorithme 4.1). La structure est très proche de l’algorithme 3.2. Dans cet algo-
rithme on conserve la mutation à côté de la recherche locale, mais on pourrait
effectuer l’un ou l’autre, ou les deux, systématiquement ou sous condition de pro-
babilité. Les possibilités sont infinies. C’est ce qui complique sans doute la mise
au point de ces méthodes.
L’intensification dans cet algorithme est produite de toute évidence par l’ap-
plication nouvelle de la recherche locale. L’opérateur de mutation assure la diver-
sification de la méthode. Par ailleurs, il existe des comportements pas toujours
souhaitables. On constate par exemple que lorsque l’on choisit un codage indirect
80
4.2 Scatter search
de la solution, il n’est pas toujours facile de trouver des opérateurs de croisement

efficaces qui permettent d’hériter des bonnes caractéristiques des parents. Une
petite étude statistique menée sur la qualité des parents et de leurs enfants tend à
prouver que la diversification pourrait être aussi amenée par l’opérateur de croi-
sement lui-même [92]. Pourtant son but initial dans l’algorithme génétique est
bien d’intensifier la recherche. Nous nous demandons alors, si la diversification
n’est pas amenée simplement par l’ajout de nouvelles solutions de départ dans
une recherche locale (un peu comme dans un multi-start, sauf que les solutions
initiales ne sont pas aléatoires, mais résultant de l’opérateur de croisement).
4.2 Scatter search
L’origine de la méthode ou du moins la première publication que l’on trouve

comportant ce nom est due à Glover [41]. Mais le travail conjoint de Fred Glover,
Manuel Laguna et Raphael Martı̀ a permis d’obtenir une méthode plus claire et
plus sophistiquée, et qui garde un caractère générique indéniable [46].
Cette méthode fait évoluer une population de solutions et s’appuie sur le
principe suivant. Une population de solutions (assez importante au départ) est
générée (en essayant de proposer des solutions diverses les unes des autres) et
chaque individu est amélioré par l’application d’une recherche locale. De cette po-
pulation on extrait un ensemble de référence (Reference set, R) contenant les meilleures
solutions de la population initiale. Ensuite ces solutions sont combin ées entre
elles (et avec les nouvelles solutions générées) puis améliorées jusqu’à ce qu’il n’y
ait plus de nouvelles solutions générées par combinaison. Ensuite une moitié de
la population est régénérée (remplacée par des solutions diverses) et le processus
recommence jusqu’à satisfaction d’un critère d’arrêt.
Un des points importants, c’est la mesure de la diversité des solutions. Cette
mesure doit être prise dans l’espace des solutions (et non dans l’espace des ob-
jectifs) et elle doit refléter la différence entre deux solutions. Ce type de mesure
sera utilisé dans la métaheuristique GA|PM de la section 4.3. Pour plus de détails
sur les mesures que l’on peut utiliser, voir la thèse de Sörensen [121] et Sörensen
[120].
La particularité de cette méthode est l’acharnement à épuiser les ressources
par combinaison. Ceci se traduit par une sorte d’exploration systématique de tout
un voisinage possible et bien sûr par un temps d’exécution parfois prohibitif.
Par la suite, on remplace une partie de la population et on recommence. Cette
technique de restart est présente dans d’autres algorithmes hybrides (voir section
4.4). Nous renvoyons vers les publications suivantes pour plus de détails sur la
méthode elle-même et sur des particularités d’implémentation [42, 43, 46].
L’algorithme 4.2 présente un squelette du scatter search. Volontairement nous
avons utilisé une notation ensembliste plus claire à nos yeux. Dans cet algorithme,
81
Algorithme 4.2 Basic scatter search

1: initialise : generate an initial improved population
2: Select a diverse subset R (Reference set)
3: while stopping criterion is not satisfied do
4: A←R
5: while A 6= ∅ do
6: combine solutions (B ← R × A)
7: improve solutions of B
8: update R (keep best solutions from R ∪ B)
9: A ← B−R
10: end while
11: Remove half of the worst solutions in R
12: Add new diverse solutions in R
13: end while
l’idée générale des auteurs est de garder à chaque itération (celles de la boucle
externe) un ensemble de référence qui soit à la fois composé de bonnes solutions
et de solutions diverses. Les facteurs d’intensification se retrouvent alors dans
l’application systématique d’une méthode de recherche locale (lignes 1 et 7) ainsi
que dans la combinaison répétée des solutions (boucle des lignes 5 à 10) et dans
la mise à jour de R (ligne 8). Le choix des sous ensembles diverses (ligne 2) et
le remplacement d’une partie de la population à chaque étape (ligne 12) sont les
facteurs de diversification du scatter search.
4.3 GA|PM
En partant des observations faites sur la plupart des métaheuristiques précé-

dentes, y compris des plus sophistiquées, plusieurs caractéristiques importantes
sont apparues à Sörensen [121]. Nous le rejoignons sur la plupart des points. En
premier lieu, une recherche locale est indispensable. De même, il est important
de mesurer (comme dans le scatter search) la diversité des solutions. Par ailleurs,
l’exploration systématique n’est pas toujours nécessaire. De ces constatations,
Sörensen [121] a proposé cette nouvelle métaheuristique appelée GA|PM (genetic
algorithm with population management – algorithme génétique avec gestion de
la population).
Le fonctionnement est assez simple et est basé sur un algorithme génétique (ici
dans sa version incrémentale). Nous supposons que nous savons comparer deux
individus entre eux et mesurer leur dissemblance. Nous pouvons donc mesurer
la similarité entre un individu et la population existante. Au départ, on génère
une population initiale de petite taille et on choisit un paramètre ∆ fixant le ni-
veau de dissemblance des solutions entre elles. Ensuite, on procède comme dans
un algorithme génétique, on choisit deux individus que l’on croise pour obtenir
82
4.3 GA|PM
deux enfants. Pour chacun on applique une recherche locale de façon à obtenir
des optima locaux. S’ils ne répondent pas au critère de diversité, on applique un
opérateur de mutation sur ces individus jusqu’à satisfaction de ce critère. Ensuite
sous condition, on les insère dans la population à la place d’un autre individu.
A chaque itération le paramètre ∆ gérant la diversité est mis à jour (suivant plu-
sieurs schémas – voir [121, 122]).
Ce qui fait sans doute que cette méthode donne de très bons résultats, c’est
qu’elle combine les avantages du scatter search et d’un algorithme mémétique.
D’un côté elle applique une recherche locale à toutes les nouvelles solutions et de
l’autre, elle maintient une population de petite taille et diversifiée. A la différence
du scatter search, la diversité est contrôlée précisément. L’autre avantage, c’est
l’évolution du paramètre de diversité ∆ (comme dans les schémas de refroidisse-
ment de température du recuit simulé) qui permet à tout moment d’augmenter
ou de réduire la diversité des individus dans la population. Pour une introduction
générale de cette méthode précisant des détails importants, consultez Sörensen et
Sevaux [122].
Algorithme 4.3 GA|PM outline

1: initialise population P
2: set population diversity parameter ∆
3: repeat
4: select : p1 and p2 from P
5: crossover : p1 ⊗ p2 → c1 , c2
6: local search : on c1 and c2
7: for each child c do
8: while d P (c) < ∆ do
9: mutate c
10: end while
11: if c satisfies conditions for addition then
12: remove solution : P ← P\b
13: add solution : P ← P ∪ c
14: end if
15: end for
16: update diversity parameter ∆
L’algorithme 4.3 présente un cadre général de l’algorithme GA|PM qui peut

sans doute être amélioré pour chacun des problèmes que l’on souhaite traiter.
Dans cet algorithme le facteur d’intensification principal est l’appel à la recherche
locale. Concernant la diversification, c’est la boucle (lignes 8 à 10) imposant un
certain niveau de diversité qui joue ce rôle. La manipulation du paramètre ∆ peut
être interprétée dans les deux sens. Si on diminue ∆, on autorise l’intégration de
solutions moins diverses, donc on intensifie la recherche, dans le cas contraire, on
83
force les individus à être de plus en plus éloignés entre eux, on diversifie donc la
recherche. Ce qu’il serait intéressant de mesurer ici, c’est le nombre d’itérations de
la boucle des lignes 8 à 10 pour voir si la recherche locale concentre les individus
dans une même région. Cela permettrait sans doute un ajustement plus précis de
∆ au cours de la recherche.
4.4 Applications
Les algorithmes mémétiques sont plus anciens et sans doute plus faciles à implé-
menter (il suffit de combiner le code d’un algorithme génétique et d’une recherche
locale). On trouve naturellement plusieurs centaines d’applications. Pour des ap-
plications intéressantes, nous renvoyons le lecteur vers http://www.densis.fee.
unicamp.br/~moscato/.
La scatter search plus récente commence à avoir les faveurs des chercheurs.
Le retard pris dans l’utilisation de la méthode provient très certainement du fait
que les papiers originaux étaient peu clairs sur certaines parties de l’algorithme
lui-même. Il existe aujourd’hui un certain nombre d’articles acceptés ou de rap-
ports de recherche qui présente des applications très intéressantes (voir http:
//www-bus.colorado.edu/faculty/laguna/scattersearch.html).
Ordonnancement à une machine

Au début de l’année 2001, nous avons commencé à travailler à la comparaison
des algorithmes mémétiques et du scatter search. Peu après, les auteurs du scat-
ter search ont proposé une série d’articles de même type (voir Campos et al.
[10], Martı́ et al. [79]). De notre côté, nous avons travaillé sur un problème d’or-
donnancement à une machine où l’objectif est la minimisation du retard total
pondéré, Sevaux et Thomin [119]. Un ensemble de n jobs (j = 1 . . . n) doit être
séquencé sur une machine unique où la préemption n’est pas autorisée. Chaque
job ne peut démarrer avant sa date de début au plus tôt, r j (release date). Le temps
d’exécution de ce job est de p j unités de temps (processing time). Une date échue,
d j (due date) et un poids, w j (weight) sont donnés pour chaque job. Si le job se
termine après sa date échue, on dit qu’il est en retard et l’on minimise la somme
des retards pondérés. Ce problème est noté 1|r j | ∑ w j Tj dans la littérature [51] et
N P-complet au sens fort. C’est sans doute l’un des problèmes à une machine les
plus difficiles. La réduction au problème sans les poids, 1|r j | ∑ Tj est N P-complet
au sens fort [35]. Si on conserve les poids, mais que les jobs sont tous disponibles
à l’instant 0, le problème 1| | ∑ w j Tj est aussi N P-complet au sens fort [74]. Il faut
arriver au problème 1| | ∑ Tj pour rencontrer un problème N P-complet au sens
ordinaire, [25].
Dans notre étude, nous souhaitons comparer un algorithme mémétique (noté
MA ici et GA dans l’article d’origine) et un scatter search (noté SS). Nous avons
84
4.4 Applications
utilisé deux jeux tests, ceux provenant de la OR-Library (ORLIB) et ceux que nous
avons créé (ODD). Les instances ODD sont tirées de Sevaux et Sörensen [111].
Pour chacun des exemples, le critère d’arrêt des algorithmes est 10 minutes ou
10000 itérations sans amélioration du meilleur individu de la population. Tous
les autres paramètres sont détaillés dans Sevaux et Thomin [119]. Les tables 5 et 6
présentent les résultats de cette étude. L’algorithme mémétique est toujours limité
par le nombre d’itérations tandis que le scatter search atteint toujours la limite de
temps. Les colonnes 2 et 3 des tables présentent donc soit le temps total utilisé
(MA) soit le nombre d’itérations avant arrêt (SS). La colonne MA = SS indique
le nombre de fois où les deux méthodes trouvent le même résultat, et la colonne
MA > SS le nombre de fois où le scatter search obtient de meilleurs résultats
(strictement) ; l’inverse ne se produit jamais.
TAB . 5 – Résultats sur ORLIB extraits de Sevaux et Thomin [119]

Instances CPU MA it. SS MA = SS MA > SS Sol.
ORLIB 40 3.89s 4836 119 6 125
ORLIB 50 8.16s 2349 0 125 125
ORLIB 100∗ 58.36s 2 42 62 54
∗Time limit = 120s
Dans la table 6 deux colonnes supplémentaires “MA best” et “SS best” in-
diquent à quel instant dans la recherche les meilleures solutions ont été trouvées.
TAB . 6 – Résultats sur ODD extraits de Sevaux et Thomin [119]

Inst. CPU MA it. SS MA best SS best MA = SS MA > SS
ODD 20∗ 13.34s 8610 0.4s 0.0s 17 0
ODD 40 5.60s 1253 0.6s 3.1s 17 3
ODD 60 21.21s 291 4.6s 19.1s 15 5
ODD 80 62.68s 99 9.2s 248.9s 9 11
ODD 100 111.16s 43 33.6s 289.0s 9 10
∗SS cannot generate 10 diverse solutions for 3 instances
Ce qu’on peut conclure de ces résultats c’est la supériorité du scatter search,

avec des paramètres identiques à l’algorithme génétique hybride. Il est impossible de
généraliser sur un seul exemple et conclure qu’on ne devrait utiliser que le scat-
ter search est faux. La difficulté d’implémentation est tout de même plus grande
pour SS et les bons résultats sont obtenus avec un plus grand délai. Par contre on
peut tout de même penser que mesurer la diversité de la population est un point
important dans la résolution.
Lors de l’écriture de Sörensen et Sevaux [122], pour prouver le bon fonctionne-
ment de GA|PM, nous avons effectué une comparaison entre le même algorithme
mémétique (MA) et le GA|PM. Dans l’article d’origine l’algorithme mémétique
est noté HGA. Les temps CPU ont été volontairement limités à une minute. Les
85
tables 7 et 8 présentent les résultats de cette étude.

Pour chaque groupe d’instances de la OR-Library (table 7), on retrouve le
nombre de solutions optimales (ou meilleures solutions connues) retrouv ées par
chaque méthode ainsi que le nombre de fois où la méthode arrive en première
position seule. L’écart moyen constaté entre la solution trouvée et la solution opti-
male (ou meilleure solution) ainsi que le nombre moyen et maximum d’it érations
est reporté.
TAB . 7 – Résultats sur ORLIB extraits de Sörensen et Sevaux [122]

ORLib40 results
Method Opt. First Avg. Gap Iterations
used sol. pos. (in %) Avg. Max
GAPM 125 4 0.000 5682 8961
MA 121 0 0.284 327763 470984
ORLib50 results
Method Opt. First Av. Gap Iterations
used sol. pos. (in %) Aver. Max
GAPM 123 11 0.002 3334 5298
MA 113 0 0.595 196024 287802
ORLib100 results
Method Best. First Av. Gap Iterations
GAPM 94 27 0.276 1030 1480
MA 85 16 2.110 34164 56521
Dans la table 8, on retrouve les mêmes informations présentées de manière

différentes (à l’exception des optima qui ne sont pas connus ici, l’écart est donc
mesuré par rapport à la meilleure solution connue).
TAB . 8 – Résultats sur ODD extraits de Sörensen et Sevaux [122]

GAPM results MA results
Set of First Avg. gap Avg. First Avg. gap Avg.
inst. pos. (in %) iter. pos. (in %) iter.
ODD20 0 0.000 32375 0 0.000 952757
ODD40 0 0.000 4142 0 0.000 258526
ODD60 6 0.088 1685 1 0.251 88106
ODD80 6 0.003 1153 1 0.083 39040
ODD100 8 0.064 844 4 0.118 19854
Global 20 0.057 8040 6 0.148 271657
86
4.4 Applications
Sur ce problème spécifique et avec les réglages que nous avons choisi (com-
muns aux deux méthodes), GA|PM prouve sa supériorité par rapport à l’algo-
rithme génétique hybride. Pour des commentaires détaillés, il faut consulter Sören-
sen et Sevaux [122].
Pour compléter cette étude, nous avons souhaité comparer le scatter search
(SS) et GA|PM. Ces tests ont été faits spécifiquement pour cette synthèse et n’ont
pas été exécutés sur le même ordinateur que les tests précédents. Bien que les
conditions d’arrêt (60 secondes) et les autres paramètres soient identiques aux
autres tests, les valeurs trouvées (notamment les optima retrouvés pour ORLIB)
ne sont pas toujours les mêmes.
Dans la table 9, on constate que bien que le temps de résolution soit court, ce
qui est en général défavorable au SS, il conserve un léger avantage. Le nombre de
fois où SS trouve une meilleure solution (strictement) que GA|PM est supérieur
pour les grandes instances (100 jobs) et les écarts mesurés à la meilleure solu-
tion sont aussi plus petits (0.011% pour SS contre 0.128% pour GA|PM). Pour le
cas particulier du Scatter Search, nous avons ajouté le nombre moyen de cros-
sover (colonne Xov.) qui donne une valeur du même ordre de grandeur que les
itérations du GA|PM.
TAB . 9 – Résultats comparatif de SS et GA|PMs sur ODD

GAPM results SS results
Set of First Avg. gap Avg. First Avg. gap Avg. Avg.
inst. pos. (in %) iter. pos. (in %) iter. Xov.
ODD20 0 0.000 33496 0 0.000 168 25509
ODD40 0 0.000 4142 0 0.000 31 5710
ODD60 0 0.057 1685 2 0.000 8 1859
ODD80 2 0.000 1153 0 0.013 3 1021
ODD100 3 0.128 844 7 0.010 1 938
Global 5 0.112 6070 9 0.011 31 5405
Dans la table 10, la tendance s’inverse et GA|PM prend légèrement le dessus

sur SS. Le nombre de solutions optimales (ou meilleures solutions) retrouv é est
toujours plus important pour GA|PM et en conséquence l’écart à la solution opti-
male (ou meilleure solution) est plus important pour SS.
On peut noter aussi que nous avons compté le nombre d’itérations de SS en
comptant les itérations de la boucle extérieure. Pour corriger cette erreur, nous
avons aussi ajouté le nombre de nouvelles solutions générées par croisement (co-
lonne Xov.). Ce nombre est du même ordre de grandeur que le nombre d’itérations
de GA|PM. En comparant les valeurs obtenues, on remarque que pour un même
temps d’exécution, le nombre d’itérations de SS est souvent plus grand.
87
TAB . 10 – Résultats comparatif de SS et GA|PM sur ORLIB

ORLib40 results
Method Opt. First Avg. Gap Iterations Xov.
used sol. pos. (in %) Avg. Max Avg.
GAPM 120 2 0.000 5772 8961
SS 118 0 0.125 44 99 15703
ORLib50 results
Method Opt. First Av. Gap Iterations Xov.
used sol. pos. (in %) Aver. Max Avg.
GAPM 120 4 0.002 3380 5298
SS 117 1 0.011 20 36 9406
ORLib100 results
Method Best. First Av. Gap Iterations Xov.
used sol. pos. (in %) Aver. Max Avg.
GAPM 94 44 0.276 1030 1480
SS 73 11 2.008 2 3 1905
Tournées de véhicules sur arcs

Le même type de comparaison est conduit sur un problème différent. Il s’agit du
Capacitated Arc Routing Problem (CARP). Dans ce problème, une flotte de véhicules
identiques partent d’un dépôt, collectent des marchandises (le long des arcs d’un
graphe définissant le réseau) et reviennent au dépôt lorsqu’ils sont remplis. L’ob-
jectif est de minimiser le trajet total parcouru par l’ensemble des véhicules. Pour
une description complète du problème lire Lacomme et al. [69]. Les meilleurs
résultats sont proposés par Lacomme et al. [68] et pour d’autres techniques, on
peut se reporter à Greistorfer [52] qui proposent un “tabu scatter search” pour
résoudre le problème.
Dans la table 11 on retrouve les résultats de cette étude sur les instances de
Golden et al. [50]. La première colonne indique les algorithmes utilisés. SMA
correspond à un algorithme mémétique simple. BMA est le meilleur algorithme
résolvant ce problème aujourd’hui (voir Lacomme et al. [68]) et c’est un algo-
rithme mémétique dans lequel une technique simple est utilisée pour disper-
ser la population. L’indicateur “nr” est présent quand les restarts n’ont pas été
utilisés (pas de redémarrage de l’algorithme avec une partie de la population
renouvelée). Pour chaque algorithme, la déviation à la meilleure borne connue
(donnée par Belenguer et Benavent [6]) est donnée en moyenne et au pire. La
colonne “LB hits” donne le nombre de fois où la borne inférieure est atteinte et
la colonne suivante le temps CPU en secondes. La colonne “Restarts” indique le
nombre d’instances nécessitant des restarts et entre parenthèses le nombre d’ins-
tances pour qui les restarts ont conduit à l’optimum. La dernière colonne indique
88
le nombre moyen de croisement effectués.
TAB . 11 – Résultats du CARP extraits de Prins et al. [93]

Algorithm Dev. LB Worst LB hits Av. time Restarts Av. Xovers
SMA 0.65 4.07 15 0.37 0 2750.5
BMA-nr 0.33 2.23 18 0.95 0 3013.1
GAPM-nr 0.24 2.23 20 0.90 0 880.9
BMA 0.17 2.23 21 4.79 5(3) 9960.2
GA|PM 0.17 2.23 21 1.59 3(1) 1968.9
SMA est très rapide mais donne des résultats loin des meilleurs. Sans les
restarts, GA|PM est bien meilleur que BMA en terme d’écart et en nombre de
solutions optimales retrouvées. Si on autorise les restarts, le même nombre de
solutions optimales est trouvé, mais GA|PM garde un avantage en temps de
résolution. Le nombre d’instances pour lesquelles les restarts sont nécessaires est
moins important. C’est donc la meilleure méthode résolvant ces instances aujour-
d’hui.
5 Compléments
Il existe bien d’autres méthodes qui auraient pu figurer dans ce document.
Nous les avons volontairement ou non écartés. D’un côté, certaines de ces mé-
thodes n’ont pas prouvé leur efficacité et d’autres peuvent se trouver être des
cas particuliers des méthodes déjà présentées. Pour toutes les méthodes que nous
avons oublié et pour les auteurs qui auraient souhaité avoir leur place dans ce do-
cument de synthèse, nous présentons nos plus plates excuses. Avant de conclure,
nous pensons qu’il est important de mettre en avant des points relatifs aux mé-
taheuristiques que nous avons abordés au cours des dernières années et qui sont
pour certains un nouveau domaine d’investigation ou pour d’autres un champ
d’application et de recherche à part entière.
Ci-dessous, nous soulevons un certain nombre de particularités et de ques-
tions qui nous paraissent intéressantes et auxquelles nous tenterons de répondre
dans la conclusion de cette synthèse.
Codage d’une solution

L’implémentation de chacune des métaheuristiques des sections précédentes peut
parfois poser des problèmes. En premier lieu, et quelque soit la méthode utilisée,
le codage de la solution est le point crucial de la réussite de l’approche. Par un
codage approprié, qu’il soit direct ou indirect, on manipule plus ou moins facile-
ment les solutions de notre problème, on est plus ou moins rapidement capable
89
5 Compléments
d’évaluer la qualité d’une solution et on peut extraire plus ou moins d’informa-

tions intéressantes sur l’espace des solutions.
Voisinage
En ce qui concerne les méthodes de recherche locale, la définition d’un bon voi-
sinage est particulièrement sensible. Il est par exemple très important de savoir
si le voisinage est connexe ou non. Et s’il ne l’est pas, comment être certain de ne
pas passer à côté de la solution optimale ? La taille du voisinage en lui-même est
un critère intéressant. Si le voisinage est petit, on va pouvoir l’explorer tout en-
tier, mais on aura besoin d’énormément d’itérations pour explorer un espace des
solutions le plus large possible. Si le voisinage est grand, le nombre d’it érations
nécessaire sera sans doute plus petit, mais le temps nécessaire pour parcourir ce
voisinage peut être exponentiel...
Opérateurs de croisement et de mutation

Pour les méthodes à population (à l’exception des colonies de fourmis), les opé-
rateurs de croisement et de mutation sont toujours difficiles à choisir. Il existe
heureusement dans beaucoup de domaines (comme par exemple en ordonnan-
cement, Caux et al. [11], Portmann [91]) des études sur les différents opérateurs
et leur performance en fonction du codage et du problème lui-même. Une ca-
ractéristique importante d’un opérateur de croisement est qu’il devrait “en prin-
cipe” produire un enfant au moins aussi bon que les parents.
Mesure de la diversité
Une autre question importante à nos yeux et indispensable pour le scatter search
et le GA|PM est la définition d’une mesure de distance entre individus. Celle-ci
doit être capable de refléter une similitude ou une différence entre deux indivi-
dus. On pourra alors en déduire la distance d’un individu à la population toute
entière. En fonction du problème, du codage des individus et de la distance re-
tenue, l’évaluation peut-être plus ou moins coûteuse en temps de calcul. Mais
parfois il peut-être bon de passer du temps pour calculer cette distance si elle
conduit au final à une meilleure diversité de la population. Certaines distances,
principalement pour les codages de type permutation, sont décrites dans Campos
et al. [10], Sörensen [120], Sörensen et Sevaux [122].
5.1 Réglages automatiques des paramètres
Passer de la théorie à la pratique est souvent difficile. L’implémentation de

toutes ces métaheuristiques pose des problèmes insoupçonnables. Un des sou-
cis, et sans doute pas le moindre, est la compétence informatique des chercheurs
opérationnels. Savoir programmer dans un langage informatique de 4e génération
90
5.2 Robustesse
(quel qu’il soit) est très important. Si on suppose que cette compétence est acquise,
alors un autre problème va surgir : le réglage des paramètres.
Prenons quelques exemples. Pour le recuit simulé, il faut trouver la température
de départ, et le schéma de variation de cette température au cours des itérations.
Pour la recherche tabou, la longueur de la liste tabou est importante. Trop grande,
celle-ci conduit à un blocage où dans le pire des cas tous les voisins d’une solution
sont tabous ; trop petite, celle-ci ne peut éviter les cyclages et la méthode parcoure
les mêmes solutions indéfiniment.
Nous proposons dans Sevaux et Thomin [115, 119] la mise en place de tech-
niques de réglages automatiques de ces paramètres. Pour le recuit simulé, notre
expérience est plutôt défavorable. Le but recherché était d’initialiser automatique-
ment la température de départ pour donner environ une chance sur deux à une
solution dégradante d’être acceptée. Malheureusement, les liens avec le schéma
de refroidissement de la température sont trop forts et ce réglage automatique
dans certains cas nous a déçu. Pour la recherche tabou, nous sommes plus enthou-
siastes. Dans la méthode présentée par Sevaux et Thomin [115], une détection au-
tomatique de cyclage est effectuée en permanence et en cas de détection, la lon-
gueur de la liste est augmentée de la longueur du cycle. Ensuite la longueur est
progressivement diminuée pour éviter de conserver des éléments non-nécessaires
qui augmentent le temps de calcul de la vérification du statut tabou. Cette méthode
s’apparente un peu à la recherche tabou réactive proposée par Battiti et Tecchiolli
[3].
5.2 Robustesse
La recherche de solutions robustes à un problème d’optimisation semble être

un challenge récent pour les chercheurs. Longtemps laissé aux statisticiens ou
stochasticiens, ce domaine est en train de voir sa popularité augmenter dans la
communauté de la recherche opérationnelle, Sevaux et Sörensen [110].
Quand le problème est déterministe c-à-d. que les données sont connues et
fiables à 100%, la plupart du temps le problème sous-jacent est un problème d’op-
timisation N P-difficile. Il n’existe alors pas d’algorithme efficace pour le r ésoudre
optimalement. Les métaheuristiques deviennent un outil de résolution privilégié
garantissant à la fois un temps de développement réduit et l’obtention d’une so-
lution de bonne qualité. Même si les métaheuristiques et particulièrement les
algorithmes génétiques sont appliqués souvent et avec succès pour traiter des
problèmes, leur application en environnement incertain est très limitée. Nous
affirmons que les métaheuristiques peuvent être très facilement adaptées aux
nécessités d’un problème stochastique. L’utilisation de métaheuristiques pour
ce type d’optimisation présente un certain nombre d’avantages (voir Sevaux et
Sörensen [111], Sevaux et al. [112]). Le besoin pour l’optimisation robuste par
métaheuristiques a été reconnu dans le livre majeur sur l’optimisation robuste
91
5 Compléments
[66], quand les auteurs énoncent page 354 ¡¡ We believe that considerable more effort
should be spent on the systematic development of [...] metaheuristic frameworks, which
with minimal adjustment effort can be applied to a large class of robust optimisation pro-
blem [...] ¿¿.
On trouve dans la littérature un nombre impressionnant d’articles utilisant
les métaheuristiques pour résoudre des problèmes d’ordonnancement. Parmi les
techniques les plus prisées, on retrouve en tête les algorithmes génétiques et la re-
cherche tabou. L’article de Reeves [96] présente une liste non-exhaustive d’appli-
cations des algorithmes génétiques dans la plupart des domaines de l’optimisa-
tion combinatoire. Pour la robustesse en ordonnancement, nous avons tent é d’ap-
porter un certain nombre de réponses en utilisant le cadre général proposé par
Sörensen [121] et que nous avons appliqué à un problème d’ordonnancement Se-
vaux et Le Quéré [108], Sevaux et al. [112]. En utilisant des particularités stochas-
tiques de notre problème, on peut intégrer ces dernières à une métaheuristique.
On obtient alors une métaheuristique “stochastique” hybride qui semble parti-
culièrement bien s’adapter au problème Fleury et al. [32].
Dans la gamme des problèmes de tournées de véhicules, nous tentons d’ap-
pliquer encore une fois le cadre général de Sörensen [121]. Ce type d’approche
semble être satisfaisant pour la résolution des différentes variantes stochastiques
du problème VRP (VRP avec temps de parcours stochastiques). Cette fois-ci nous
nous appuyons sur un algorithme de type GA|PM que nous avons adapté pour
obtenir des solutions robustes, Sörensen et Sevaux [123]. Cette technique, même
si elle apporte des solutions satisfaisantes, doit être couplée avec une évaluation
stochastique assez précise de la robustesse d’une solution comme dans Fleury
et al. [31].
5.3 Optimisation multiobjectif
Plus brièvement, nous souhaitons mettre en avant un domaine nouveau pour

nous, mais dont la communauté, plutôt constituée d’informaticiens, s’est ouverte
au monde de la recherche opérationnelle récemment. C’est sans doute grâce aux
efforts permanents de plusieurs personnes, comme Xavier Gandibleux, qui n’ont
pas hésité pas à s’investir énormément sur cette thématique fort intéressante (voir
Gandibleux [34]).
La résolution de tournées de véhicules sur arcs est un problème dont les appli-
cations sont particulièrement importantes pour les industriels. La première appli-
cation que nous souhaitons résoudre est la collecte de déchets en milieu urbain.
Après plusieurs conversations avec des industriels, il est apparu que l’objectif
classiquement traité (la minimisation de la durée totale des tournées) n’était pas
toujours l’objectif prioritaire. Dans certains cas, obtenir à la fois des durées to-
tales minimales et un équilibre entre les tournées s’avérait plus intéressant dans
la réalité. Nous avons traité ce problème en appliquant une variante de NSGA-II
92
5.4 Optimisation continue
(un algorithme génétique dédié à la résolution de problèmes multiobjectifs) et les

résultats obtenus sont très satisfaisants (voir Lacomme et al. [72, 73]).
5.4 Optimisation continue
La dernière application que nous souhaitons présenter sort totalement du

cadre de l’optimisation combinatoire. La conception assistée par ordinateur en
“design” automobile est partout.
Le procédé de création d’un véhicule prototype est très particulier. Dans un
premier temps, un modèle est conçu sur ordinateur. A partir de ce modèle, le desi-
gner va fabriquer une maquette qu’il va re-modeler manuellement jusqu’à obtenir
la forme désirée. Cette étape est indispensable car il est impossible au designer
de se rendre compte du résultat sur un écran d’ordinateur. Une fois la maquette
obtenue dans sa version définitive, par un procédé de reverse engineering, on me-
sure des quantités de points sur la maquette et on approxime ces points par de
nouvelles courbes.
Trouver les paramètres de ces courbes est extrêmement difficile et les réglages
manuels sont souvent impossibles à reproduire plusieurs fois à l’identique. Avec
Y. Mineur, nous avons mis au point un algorithme génétique continu (basé sur
les travaux de Chelouah et Siarry [13]) qui permet d’offrir une aide en CAO pour
l’obtention de ces paramètres. L’ensemble des résultats est présenté dans Mineur
et Sevaux [83] et dans Sevaux et Mineur [109].
6 Conclusions
Après ce tour d’horizon assez vaste sur les métaheuristiques et un regard sur
les différentes applications que nous avons pu traiter, sommes-nous aujourd’hui
capable de définir une métaheuristique universelle ? Probablement pas ! Serait-ce
alors un constat d’échec ? Pas plus !
Ce que nous allons proposer c’est plutôt des indications sur les caractéristiques
d’une bonne métaheuristique à notre sens. Voici donc des éléments importants à
nos yeux :
Population Les méthodes à population ont un avantage certain, celui du pa-
rallélisme intrinsèque qu’elles induisent. Travailler sur un ensemble de so-
lutions qui partagent des informations, c’est explorer un plus grand es-
pace des solutions. De plus, l’échange d’informations entre les individus
de différentes régions de l’espace des solutions est souvent bénéfique.
Codage Le codage retenu pour la représentation de la solution est un point clé
comme on a pu le voir. Il peut être direct ou indirect (phénotypique ou
génotypique comme diraient les promoteurs des algorithmes génétiques).
93
6 Conclusions
De notre point de vue et expérience, un codage indirect est souvent plus fa-
cile à manipuler (du point de vue de la structure informatique des données)
mais nécessite un algorithme d’évaluation de la solution. C’est au concep-
teur de la méthode de trouver un juste équilibre entre les deux. D’expérience,
un algorithme d’évaluation d’un individu de la population en O(n 2 ) reste
raisonnable. A noter que l’on obtient de très bons résultats quand l’évaluation
peut être optimale.
Coopération Les individus doivent être capables d’échanger des informations
entre eux par croisement par exemple. Le croisement est souvent fait en
prenant les individus deux à deux, mais comme le propose le scatter search,
ce n’est pas une obligation et plusieurs individus peuvent coopérer pour
fournir de nouvelles solutions. L’idéal serait de pouvoir récupérer systé-
matiquement les bonnes caractéristiques des parents pour les transmettre
aux enfants. Mais quelles sont ces bonnes caractéristiques, souvent on ne le
sait pas. Un point que soulignent Hertz et Widmer [60] et sur lequel nous
sommes d’accord, c’est que si l’on croise deux parents identiques, il semble
raisonnable d’obtenir un enfant qui soit identique aux deux parents. Cette
propriété n’est malheureusement pas toujours respectée.
Recherche locale Ajouter une recherche locale est aujourd’hui indispensable. Les
méthodes qui n’en sont pas dotées ne sont pas compétitives en terme de
qualité de solutions proposées. Par contre, si le codage retenu est indirect,
la recherche locale doit se faire en tenant compte de la solution elle-même.
Un bon exemple pourrait être celui de l’algorithme mémétique de Lacomme
et al. [69]. Le codage retenu est une permutation, mais la recherche locale est
effectuée en prenant des voisinages classiques en “tournées de véhicules”.
Un mauvais exemple est celui de l’algorithme génétique proposé dans [119]
où le codage d’une solution est une permutation (comme souvent en or-
donnancement) et la recherche locale se fait en utilisant le voisinage gene-
ral pairwise interchange (GPI) qui consiste à échanger deux éléments de la
permutation sans intégrer des propriétés importantes du problème. Cette
erreur n’a pas été reproduite dans [115].
Diversité contrôlée Mesurer la diversité entre les solutions est particulièrement
important. Quand cela est intégré à un mécanisme de contrôle de la diver-
sité, la métaheuristique peut se concentrer sur les individus divers et ne pas
effectuer d’exploration inutile. Le contrôle de la diversité est un facteur tout
aussi important. Il permet en fonction de l’ensemble des individus de la
population de plus ou moins diversifier et de plus ou moins intensifier la
recherche.
Restarts Les observations des techniques de “restart” implémentées, et notam-
ment celles que nous avons testé dans Prins et al. [93], montrent qu’il est
parfois envisageable de re-générer une partie de la population. En effet,
après plusieurs générations, une partie de la population tend à être très per-
formante, tandis qu’une autre partie se retrouve sans potentiel intéressant
94
pour des croisements futurs. En remplaçant une partie de cette population
(la moins bonne de préférence) on donne un sang neuf à la population et on
peut ainsi parfois améliorer les meilleures solutions.
Aléatoire Les plus grandes découvertes du monde l’ont été pour certaines d’entre
elles par le fait du hasard. Se priver de tout facteur aléatoire implique qu’il
faut remplacer cette part de chance par une énumération complète, que ce
soit au niveau de la sélection des individus ou par exemple des points de
croisement d’un opérateur de croisement. Par certains côtés, c’est ce que
fait le scatter search, mais au prix toujours d’un temps de calcul souvent
important.
Si ces caractéristiques sont les bonnes, alors l’algorithme GA|PM y répond
quasiment en totalité et c’est désormais le seul que nous devrions utiliser. Non,
surtout pas ! Ce serait sans doute une grave erreur. Ce que nous n’avons pas pris
en compte dans les caractéristiques ci-dessus, c’est 1) le problème lui-même et 2)
le but de la résolution.
– Le problème en lui-même est un point fondamental dans la résolution. Il
peut être plus ou moins simple à coder, à comprendre et à résoudre. Il peut
exister des spécificités à ce problème, comme des cas particuliers polyno-
miaux sur lesquels s’appuyer. Par exemple, pour un problème d’ordonnan-
cement juste-à-temps, si on connaı̂t les jobs en avance et ceux en retard, sous
certaines conditions on peut évaluer optimalement cette configuration. Du
problème et de ses caractéristiques, on va pouvoir choisir le codage le plus
approprié, l’évaluation la meilleure, etc.
– Le but de la résolution est primordial et souvent ignoré. Dans le contexte
académique, on cherche à fournir le plus souvent la meilleure solution au
prix d’une implémentation parfois difficile avec des temps de calculs sou-
vent longs et le résultat est l’amélioration des meilleures solutions mais des
fois de quelques pour cents uniquement. Si on s’adresse à un industriel, la
notion de solution optimale lui échappe souvent ou lui est indifférente. Il a
besoin de savoir si l’implémentation va être robuste, combien de temps on
va passer dessus, bref, si l’investissement en vaut la peine. Est-ce nécessaire
de trouver la solution optimale à un problème industriel auparavant résolu
manuellement et dont une solution obtenue par simple heuristique améliore
un rendement de plus de 50% ? Martin Grötschel lors du Euro Winter Ins-
titute en Mars 2000 rapportait que pour un problème de transport de per-
sonnes handicapés (dial-a-ride problem) il avait trouvé la solution optimale
en quelques semaines de travail mais avait mis plus de cinq ans à la faire
appliquer pour ne pas modifier en profondeur le travail des conducteurs du
jour au lendemain.
De toute notre expérience, le seul impératif qui ressort c’est qu’il faut es-
sayer. En conférence, dans les présentations sur les métaheuristiques, les ques-
tions simples comme “Mais pourquoi n’avez-vous pas utilisé la méthode X plutôt
que la méthode Y ?” ont heureusement pratiquement disparu. Un seul mot d’ordre
95
7 Perspectives de recherche
à cela, il faut essayer. On ne peut assurément pas répondre à une question de-
mandant le résultat d’une autre métaheuristique sans l’avoir implémenté et testé
auparavant.
Un bon indicateur des perspectives de recherche est le nombre de publications
en cours ou en préparation [16, 108, 109, 112, 115, 122]. Pour certaines d’entre
elles, le sujet est bien abouti et sera sans doute terminé après la publication d’un
article en revue. Pour d’autres, il y a encore beaucoup à faire.
Nous allons ici lister les thèmes en cours d’étude ou qui présentent un intérêt
particulier et qui seront privilégiés dans l’avenir.
Résolution de problèmes multiobjectifs

Ce nouveau type de problèmes, abordé sur le cas particulier du CARP [70,
72], a permis de montrer les lacunes des méthodes existantes. On peut par
exemple voir que beaucoup d’algorithmes génétiques multiobjectifs ont des
fonctionnements simplistes et que ce qui existe pour les problèmes mono-
objectif pourrait être transposé ici. Un exemple simple, les conditions d’arrêt
des algorithmes multiobjectifs ne sont pas encore ou sont peu traitées.
Développement de méthodes robustes
Les problèmes rencontrés en industrie ont rarement des données fiables.
Pourtant, toutes les méthodes d’optimisation travaillent sur des données
fixées et connues à l’avance. Les premiers travaux [108, 110, 112] en ordon-
nancement sont prometteurs mais une meilleure intégration des données et
particulièrement au niveau statistique permettrait de proposer des méthodes
plus fiables, et moins consommatrices de temps de calcul. Ce sujet est déjà
bien avancé Fleury et al. [32].
Problèmes dans les graphes
Avec C. Prins et P. Lacomme, nous avons mis en projet de rénover les ap-
proches algorithmiques sur les graphes en proposant d’importantes biblio-
thèques “open source” orientées objet permettant d’utiliser ces algorithmes
dans la résolution de problèmes plus complexes. Ce travail est partiellement
accompli dans Lacomme et al. [71]. Une version C++ utilisant la STL, plus
générique, est aussi prévue en collaboration avec P. Thomin. Elle fera l’objet
d’un second tome de ce livre.
Ce thème de prédilection tient toujours une place importante. On a noté
que dans certains cas, l’espace des solutions est trop grand pour être traité
avec un logiciel commercial. Une technique de coupes ou de décomposition
pourrait alors être envisagée. Mais pourquoi ne pas utiliser la propagation
96
de contraintes pour réduire cet espace de solutions ? De plus, l’intégration
de métaheuristiques permettrait aussi de guider les algorithmes de pro-
grammation linéaire vers des solutions prometteuses et pourrait accélérer
la recherche. La collaboration avec C. Oǧuz sur les problèmes de type flow-
shop hybrides nous permettra de mieux cerner ces possibilités prochaine-
ment.
Mais en plus des thèmes ci-dessus, il y a deux autres domaines qui retiennent
plus particulièrement notre attention.
En premier lieu, la collaboration avec Alcatel que nous menons en ce mo-
ment (conjointement avec Xavier Gandibleux) nous a entraı̂né sur un terrain par-
ticulièrement intéressant. Le sujet est l’optimisation du trafic dans les réseaux de
télécommunications. Une clause de confidentialité nous engage à ne pas dévoiler
le cœur du travail que nous menons, mais c’est un domaine nouveau pour nous et
particulièrement intéressant. Le contrat d’un an que nous avons signé avec notre
partenaire industriel sera très probablement renouvelé et la charge de travail étant
importante, il me parait indispensable d’engager une thèse sur ce domaine. Les
moyens financiers obtenus lors de cette collaboration sont importants et pour-
raient permettre de proposer à un jeune chercheur brillant un contrat de thèse
pour réaliser cette étude avec nous.
Alcatel est un des organismes à la pointe du progrès dans le domaine étudié,
il entend le rester et sait qu’il a besoin d’étendre ses collaborations avec le monde
académique. Le soucis de confidentialité d’une telle entreprise est légitime. On
ne peut défendre des intérêts industriels forts sans se protéger de la concurrence.
Pourtant, Alcatel connaı̂t aussi les nécessités et les impératifs de publication de
notre communauté. Ainsi, lors d’autres contrats, Xavier Gandibleux a pu après
dépôt d’un brevet dont il est co-auteur avec Alcatel publier le résultat de ses tra-
vaux. Engager une thèse sur un tel sujet, même sensible, n’est donc pas incom-
patible avec les devoirs d’un encadrant vis à vis de son doctorant (en terme de
communications et de publications).
Le dernier point qui nous anime et qui constituerait un prolongement naturel
de cette synthèse concerne les métaheuristiques évidemment.
Dans la suite logique de cette composition, et en collaboration avec C. Prins
et K. Sörensen, nous allons préparer un ouvrage recensant les principales méta-
heuristiques comme ici, mais en mettant l’accent sur l’aspect implémentation qui
nous parait indispensable et indissociable des méthodes elles-mêmes. A l’issue
de ce travail, certains problèmes ou particularités des métaheuristiques devraient
nous apparaı̂tre plus clairement encore.
En faisant cet inventaire ici, nous avons pu mettre en avant ce que pourraient
être les bonnes caractéristiques d’une métaheuristique. De notre point de vue, il
reste encore beaucoup de chemin à parcourir avant d’être en mesure de proposer
une métaheuristique universelle. Pourtant, c’est notre but aujourd’hui : propo-
97
ser une méthode suffisamment générique pour résoudre le plus grand nombre
de problèmes et pas trop générique pour que les paramètres à régler ne soient
pas trop nombreux avant que l’on puisse atteindre un degré d’efficacité suffisant.
Un deuxième pas à franchir ensuite sera de proposer, en plus du cadre général
d’une métaheuristique, un code source générique et en accès libre et sous licence
publique, qui permettra à tous d’utiliser les métaheuristiques sans avoir à refaire
des implémentations lourdes et qui consomment beaucoup de temps et d’énergie.
Réaliser tout ceci ne pourra se faire sans l’aide de plusieurs participants et notam-
ment des jeunes chercheurs dont la curiosité et l’esprit vif sont les meilleurs atouts
pour une plein réussite.
Ce projet important que nous suggérons ne va pas balayer tous les bons résul-
tats de ces dernières années sur le sujet, mais au contraire s’appuiera dessus pour
essayer de proposer une vision unifiée des métaheuristiques. Si ce but, peut-être
idyllique, n’est pas facile à atteindre, au moins aurons-nous apporté une brique
supplémentaire à l’édifice...
98
Références
[1] E.H.L. Aarts et J. Korst. Simulated Annealing and Boltzmann Machines. John
Wiley, Chichester, 1989.
[2] J.T. Alander. An indexed bibliography of genetic algorithm implemen-

tations. Technical Report 94-1-IMPLE, Dept. of Information technology
and production economics, University of Vaasa, Finland, July, 23 1999.
ftp.uwasa.fi.
[3] R. Battiti et G. Tecchiolli. The reactive tabu search. ORSA Journal on Compu-
ting, 6 :126–140, 1994.
[4] E.B. Baum. Iterated descent : A better algorithm for local search in combi-
natorial optimization problems. Technical report, Caltech, Pasadena, CA,
1986.
[5] J. Baxter. Local optima avoidance in depot location. Journal of the Operational
Research Society, 32 :815–819, 1981.
[6] J.M. Belenguer et E. Benavent. A cutting plane algorithm for the Capacita-
ted Arc Routing Problem. Computers and Operations Research, 30(5) :705–728,
2003.
[7] P. Beullens, L. Muyldermans, D. Cattrysse, et D. Van Oudheusden. A gui-

ded local search heuristic for the capacitated arc routing problem. European
Journal of Operational Research, 147 :629–643, 2003.
[8] C. Blum et A. Roli. Metaheuristics in combinatorial optimization : Over-

view and conceptual comparison. ACM Computing surveys, 35(3) :268–308,
2003.
[9] E. Bonomi et J.L. Lutton. The N-city travelling salesman problem, statisti-
cal mechanics and the metropolis algorithm. SIAM Review, 25(4) :551–568,
1984.
[10] V. Campos, M. Laguna, et R. Martı́. Context-independent scatter and tabu

search for permutation problems, Unpublished manuscript, 2002. URL
http://leeds.colorado.edu/Faculty/Laguna/publications.htm.
[11] C. Caux, H. Pierreval, et M.C. Portmann. Les algorithmes génétiques et leur

application aux problèmes d’ordonnancement. In Actes des Journées d’étude
Ordonnancement et Entreprise : Applications Concrètes et Outils pour le Futur,
pages 5–45, Toulouse, 1994. CNRS.
[12] R. Cerf. Une théorie asymptotique des algorithmes génétiques. PhD thesis, Uni-
versité de Montpellier II, France, 1994.
99
Références
[13] R. Chelouah et P. Siarry. A continuous genetic agorithm designed for the

global optimization of multimodal functions. Journal of Heuristics, 6 :191–
213, 2000.
[14] N.E. Collins, R.W. Eglese, et B.L. Golden. Simulated annealing : An annota-
ted bibliography. American Journal of Mathematical and Management Sciences,
8 :209–307, 1988.
[15] D. Connolly. An improved annealing scheme for the qap. European Journal
of Operational Research, 46 :93–100, 1990.
[16] S. Dauzère-Pérès et M. Sevaux. An efficient formulation for minimizing
the number of late jobs in single-machine scheduling. Technical Report
98/9/AUTO, Ecole des Mines de Nantes, July 1998. Submitted.
[17] S. Dauzère-Pérès et M. Sevaux. Using lagrangean relaxation to minimize
2003.
[18] K.A. De Jong. Genetic algorithms : A 25 year perspective. In J.M. Murata,
R.J. Marks II, et C.J. Robinson, editors, Computational intelligence-imitating
life, pages 125–134. IEEE Press, New York, 1994.
[19] M.L. den Besten, T. Stützle, et M. Dorigo. Design of iterated local search
algorithms : An example application to the single machine total weighted
tardiness problem. In Proceedings of EvoStim’01, Lecture Notes in Computer
Science, pages 441–452, Springer, Berlin, 2001.
[20] D. De Werra et A. Hertz. Tabu search techniques : A tutorial and an appli-
cation to neural networks. OR Spektrum, 11 :131–141, 1989.
[21] M. Dorigo. Optimization, Learning and Natural Algorithms. PhD thesis, Poli-
tecnico di Milano, Italy, 1992.
[22] M. Dorigo et G. Di Caro. The ant colony optimization meta-heuristic. In
D. Corne, M. Dorigo, et F. Glover, editors, New Ideas in Optimization, pages
11–32, London, 1999. McGraw Hill.
[23] M. Dorigo, G. Di Caro, et L.M. Gambardella. Ant algorithms for discrete
optimization. Artificial Life, 5 :137–172, 1999.
[24] M. Dorigo, V. Maniezzo, et A. Colorni. Positive feedback as a search stra-
tegy. Technical Report 91-016, Politecnico di Milano, Italy, 1991.
[25] J. Du et J.Y.T. Leung. Minimizing total tardiness on one machine is N P-
hard. Mathematics of Operations Research, 15(3) :483–495, 1990.
[26] T.A. Feo et M.G.C. Resende. A probabilistic heuristic for a computationally
difficult set covering problem. Operations Research Letters, 8 :67–71, 1989.
100
Références
[27] T.A. Feo et M.G.C. Resende. Greedy randomized adaptive search proce-
dures. Journal of Global Optimization, 6 :109–133, 1995.
[28] P. Festa. Greedy randomized adaptative search procedures. AIRO News, 7

(4) :7–11, 2003.
[29] P. Festa et M. Resende. GRASP : An annotated bibliography. In C. Ribeiro et

P. Hansen, editors, Essays and surveys in metaheuristics, pages 3–21. Kluwer
Academic Publishers, Boston, 2002.
[30] C. Fleurent et J.A. Ferland. Algorithmes génétiques hybrides pour l’optimi-

sation combinatoire. RAIRO Recherche Opérationnelle, 30(4) :373–398, 1996.
[31] G. Fleury, P. Lacomme, et C. Prins. Evolutionary algorithms for stochas-

tic arc routing problems. In G. Raidl, S. Cagnoni, J. Branke, D.W. Corne,
R. Drechsler, Y. Jin, C.G. Johnson, P. Machado, E. Marchiori, F. Rothlauf,
G.D. Smith, et G. Squillero, editors, Applications of Evolutionary Computing,
Proceedings of EvoWorkshops 2004, volume 3005, pages 501–512, Coimbra,
Portugal, 5-7 April 2004. Springer-Verlag.
[32] G. Fleury, Lacomme P, et M. Sevaux. Stochastic maintenance scheduling

problem. In Proceedings of ninth international conference on project management
and scheduling, PMS 2004, pages 405–409, Nancy, France, 26-28 April 2004.
[33] C. Gagne, M. Gravel, et W.L. Price. A new hybrid tabu-vns metaheuris-

tic for solving multiple objective scheduling problem. In Proceedings of 5th
Metaheuristic International Conference, MIC 2003, pages 22.1–7, Kyoto, Japan,
25-28 August 2003.
[34] X. Gandibleux. Optimisation multiobjectif : problèmes combinatoire,

résolutions exactes et approchées, applications. Habilitation à diriger les
recherches, Université Valenciennes, France, 12 December 2003.
[35] M.R. Garey et D.S. Johnson. Computers and Intractability : A Guide to Theory
of NP-Completeness. Freeman, San Franscisco, 1979.
[36] F. Glover. Heuristics for integer programming using surrogate constraints.

Decision Sciences, 8 :156–166, 1977.
[37] F. Glover. Future paths for integer programming and links to artificial in-
telligence. Computers and Operations Research, 13 :533–549, 1986.
[38] F. Glover. Tabu search – part I. ORSA Journal on Computing, 1 :190–206, 1989.
[39] F. Glover. Tabu search – part II. ORSA Journal on Computing, 2 :4–32, 1990.
[40] F. Glover. Tabu search : A tutorial. Interface, 20(1) :74–94, 1990. Special issue
on the practice of mathematical programming.
101
Références
[41] F. Glover. Genetic algorithms and scatter search - unsuspected potentials.

Statistics and Computing, 4 :131–140, 1994.
[42] F. Glover. A Template for Scatter Search and Path Relinking, volume 1363 of
Lecture Notes in Computer Science, pages 1–53. Springer, Springer, Berlin,
1997.
[43] F. Glover. A template for scatter search and path relinking. In J.-K. Hao,
E. Lutton, E. Ronald, M. Schoenauer, et D. Snyers, editors, Artificial Evolu-
tion, volume 1363 of Lecture Notes in Computer Science, pages 13–54, Sprin-
ger, Berlin, 1998. Springer.
[44] F. Glover et S. Hanafi. Tabu search and finite convergence. Discrete Applied
Mathematics, 119 :3–36, 2002.
[45] F. Glover et M. Laguna. Tabu Search. Kluwer, Boston, 1999.
[46] F. Glover, M. Laguna, et R. Martı̀. Fundamentals of scatter search and path

relinking. Control and Cybernetics, 39(3) :653–684, 2000.
[47] F. Glover, M. Laguna, E. Taillard, et D. De Werra, editors. Tabu search, vo-

lume 41. Annals of Operations Research, 1993.
[48] D. Goldberg. Genetic algorithms with sharing for multimodal function op-
timization. In Proceedings of the Second International Conference on Genetic
Algorithms, pages 41–49, 1987.
[49] D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Lear-
ning. Addison Wesley, 1989.
[50] B.L. Golden, J.S. DeArmon, et E.K. Baker. Computational experiments with
algorithms for a class of routing problems. Computers and Operations Re-
search, 10(1) :47–59, 1983.
[51] R.L. Graham, E.L. Lawler, J.K. Lenstra, et A.H.G. Rinnooy Kan. Optimi-
zation and approximation in deterministic sequencing and scheduling : A
survey. Annals of Discrete Mathematics, 5 :287–326, 1979.
[52] P. Greistorfer. A tabu-scatter search metaheuristic for the arc routing pro-
blem. Computers and Industrial Engineering, 44(2) :249–266, 2003.
[53] C. Guéret, C. Prins, et M. Sevaux. Programmation linéaire. Eyrolles, 2000.

[54] C. Guéret, C. Prins, et M. Sevaux. Applications of optimisation with Xpress-

MP. Dash optimization, 2002. ISBN 0-9543503-0-8, Translated by S. Heipke.
[55] B. Hajek. Cooling schedules for optimal annealing. Mathematics of Opera-

tions Research, 13 :311–329, 1988.
102
Références
[56] P. Hansen. The steepest ascent mildest descent heuristic for combinatorial
programming. In Congress on Numerical Methods in Combinatorial Optimiza-
tion, Capri, Italy, 1986.
[57] P. Hansen et N. Mladenović. An introduction to variable neighbourhood
search. In S. Voß, S. Martello, I.H. Osman, et C. Roucairol, editors, Meta-
Heuristics : Advances and Trends in Local Search Paradigms for Optimization,
pages 433–458, Kluwer, Boston, 1999.
[58] R.L. Haupt et S.E. Haupt. Practical genetic algorithm. John Wiley & Sons,
New York, 1998.
[59] A. Hertz, E. Taillard, et D. de Werra. Local search in combinatorial optimization,
chapter A tutorial on tabu search, pages 121–136. J. Wiley & Sons Ltd, New
York, 1997.
[60] A. Hertz et M. Widmer. Guidelines for the use of meta-heuristics in combi-
natorial optimization. European Journal of Operational Research, 151 :247–252,
2003.
[61] J.H. Holland. Adaptation in natural and artificial systems. Technical report,
University of Michigan, Ann Arbor, 1975.
[62] J.H. Holland, editor. Adaptation in natural and artificial systems : An introduc-
tory analysis with applications to biology, control, and artificial intelligence. MIT
Press/Bradford books, Cambridge, MA, 1992. 2nd edition.
[63] D. Johnson, C.H. Papadimitriou, et M. Yannakakis. How easy is local
search ? Journal of conputer and system sciences, 17 :79–100, 1988.
[64] S. Kirkpatrick, C.D. Gelatt, et M.P. Vecchi. Optimization by simulated an-
nealing. Science, 220 :671–680, 1983.
[65] C. Koulamas, S.R. Antony, et R. Jaen. A survey of simulated annealing
applications to operations research problems. Omega, 22 :41–56, 1994.
[66] P. Kouvelis et G. Yu. Robust Discrete Optimisation and its Applications, vo-
lume 14 of Nonconvex Optimization and its Applications. Kluwer Academic
Publishers, Dordrecht, 1997.
[67] J.R. Koza. Genetic Programming. MIT Press, Cambridge, MA, 1992.
[68] P. Lacomme, C. Prins, et W. Ramdane-Chérif. Competitive memetic algo-
rithms for arc routing problems. Annals of Operations Research, 2004. To
appear.
[69] P. Lacomme, C. Prins, et W. Ramdane-Chérif. A genetic algorithm for the
Capacitated Arc Routing Problem and its extensions. In E.J.W. Boers et al.,
editor, Applications of evolutionnary computing–EvoWorkshops 2001, Lecture
Notes in Computer Science 2037, pages 473–483. Springer, 2001.
103
Références
[70] P. Lacomme, C. Prins, et M. Sevaux. Multiobjective capacitated arc rou-

ting problem. In 2nd Joint EU/ME workshop with the french PM2O group
(MOMH Workshop) on multi-objective metaheuristics, Paris, France, 4-5 No-
vember 2002.
[71] P. Lacomme, C. Prins, et M. Sevaux. Algorithmes de graphes. Eyrolles, 2003.

[72] P. Lacomme, C. Prins, et M. Sevaux. Mutliple objective capacitated arc rou-

ting problem. In Proceedings of 2nd International Conference on Evolutionary
Multi-Criterion Optimization, EMO’2003, pages 550–564, Faro, Portugal, 8-11
April 2003. LNCS 2632.
[73] P. Lacomme, C. Prins, et M. Sevaux. A genetic algorithm for a bi-objective

capacitated arc routing problem. Computers and Operations Research, 2004.
In Press.
[74] E.L. Lawler. A psedopolynomial time algorithm for sequencing jobs to mi-
nimize total atrdiness. Annas of discrete mathematics, 1 :331–342, 1977.
[75] J.K. Lenstra, A.H.G. Rinnooy Kan, et P. Brucker. Complexity of machine

scheduling problems. Annals of Discrete Mathematics, 1 :343–362, 1977.
[76] H.R. Lourenço, O. Martin, et T. Stützle. Iterated local search. Technical

report, Technical University of Darmstadt, Germany, 2000. Preprint.
[77] H.R. Lourenço, O. Martin, et T. Stützle. Iterated local search. In F. Glover et

G. Kochenberger, editors, Handbook of Metaheuristics, pages 321–353. Kluwer
Academic Publishers, Nowell, MA, 2002.
[78] S. Lundy et A. Mees. Convergence of an annealing algorithm. Mathematical

programming, 34 :111–124, 1986.
[79] R. Martı́, M. Laguna, et V. Campos. Scatter search vs. genetic algorithms :

An experimental evaluation with permutation problems. In C. Rego et
B. Alidaee, editors, Adaptive Memory and Evolution : Tabu Search and Scatter
Search, 2002. To appear.
[80] O. Martin, S.W. Otto, et E.W. Felten. Large-step Markov chains for the tra-
veling salesman problem. Complex Systems, 5 :299–326, 1991.
[81] N. Metropolis, A. Rosenbluth, M. Rosenbluth, A. Teller, et E. Teller. Equa-

tion of state calculations by fast computing machines. Journal of Chemical
Physics, 21 :1087–1092, 1953.
[82] Z. Michalewicz. Genetic Algorithms + Data Structures = Evolution Programs.

Springer, Berlin, 1999.
104
Références
[83] Y. Mineur et M. Sevaux. Curve fitting for styling application by genetic

algorithm. In 3rd Joint EU/ME workshop with the university of Antwerp on Real-
life application of metaheuritics, Antwerp, Belgium, 18-19 December 2003.
[84] M. Mitchel. An introduction to genetic algorithms. MIT Press, Cambridge,

MA, 1998.
[85] N. Mladenović et P. Hansen. Variable neighbourhood decomposition

search. Computers and Operations Research, 24 :1097–1100, 1997.
[86] P. Moscato. On evolution, search, optimization, genetic algorithms and

martial arts : Towards memetic algorithms. Technical Report C3P 826, Cal-
tech Concurrent Computation Program, 1989.
[87] P. Moscato. Memetic algorithms : A short introduction. In D. Corne qnd

M. Dorigo et F. Glover, editors, New ideas in optimization, pages 219–234.
McGraw-Hill, New York, 1999.
[88] C.H. Papadimitriou et K. Steiglitz. Combinatorial optimization – algorithms

and complexity. Prentice-Hall, Upper Saddle River, New Jersey, 1982.
[89] L. Paquete et T. Stützle. An experimental investigation of iterated local

search for coloring graphs. In S. Cagnoni, J. Gottlieb, E. Hart, M. Midden-
dorf, et G. Raidl, editors, Applications of Evolutionary Computing, Proceedings
of EvoWorkshops2002 : EvoCOP, EvoIASP, EvoSTim, volume 2279, pages 121–
130, Kinsale, Ireland, 2002. Springer-Verlag.
[90] M. Pirlot et R.V. Vidal. Simulated annealing : A tutorial. Control & Cyberne-
tics, 25 :9–31, 1996.
[91] M.-C. Portmann. Genetic algorithm and scheduling : a state of the art and
some propositions. In Proceedings of the workshop on production planning and
control, pages I–XIV, Mons, Belgium, 1996.
[92] M.-C. Portmann. The quality of offsprings after crossover. Private commu-
nication, 2000.
[93] C. Prins, M. Sevaux, et K. Sörensen. A genetic algorithm with population

management (GA|PM) for the carp. In Proceedings of 5th trienal symposium
on transportation analysis, TRISTAN V, Le Gosier, Guadeloupe, France, 13-18
June 2004. To appear.
[94] C.R. Reeves. Modern heuristics techniques for combinatorial problems. John
Wiley & Sons, New York, 1993.
[95] C.R. Reeves. A genetic algorithm for flowshop sequencing. Computers and
Operations Research, 22 :5–13, 1995.
105
Références
[96] C.R. Reeves. Genetic algorithms for the operations researcher. INFORMS
Journal on Computing, 9 :231–250, 1997.
[97] H.-P. Schwefel. Evolution strategies : A family of non-linear optimization

techniques based on imitating some principles of organic evolution. Annals
of Operations Research, 1 :165–167, 1984.
[98] M. Sevaux. Les méthodes de recherche à voisinage. In Groupe de Recherche

en Productique – Organisation et Gesion de Production, Toulouse, France, 8-9
November 2001.
[99] M. Sevaux. Les méthodes de recherche à population. In Groupe de Recherche

en Productique - Organisation et Gestion de Production, Tarbes, France, 24-25
October 2002.
[100] M. Sevaux. Les méthodes de recherche à voisinage. Séminaire invité, Fa-

culté Polytechnique de Mons, groupe Image (in French), 6 March 2002.
[101] M. Sevaux. Metaheuristics : a quick overview. Invited seminar, University

of Antwerp, Faculty of Applied Economic Sciences, 7 February 2003.
[102] M. Sevaux. Population metaheuristics. Invited seminar, The Hong-Kong

Polytechnic University, Department of Management, 14 March 2003.
[103] M. Sevaux et S. Dauzère-Pérès. Building a genetic algorithm for a single

machine scheduling problem. In Proceedings of the 18th EURO Winter Insti-
tute, ESWI XVIII, Lac Noir, Switzerland, 4-18 March 2000.
[104] M. Sevaux et S. Dauzère-Pérès. A genetic algorithm to minimize the weigh-

ted number of late jobs on a single machine. In Proceedings of 7th Internatio-
nal Workshop on Project Management and Scheduling, PMS’2000, pages 242–
244, Osnabrük, Germany, 17-19 April 2000.
[105] M. Sevaux et S. Dauzère-Pérès. Genetic algorithms to minimize the weigh-

ted number of late jobs on a single machine. Technical Report LAMIH/SP-
2000-51, Laboratory for Automation Mechanical engineering Information
sciences and Human-machine systems, 2000.
[106] M. Sevaux et S. Dauzère-Pérès. Un algorithme génétique pour minimiser le

nombre pondéré de jobs en retard sur une machine. In 3ième conférence na-
tionale de la société française de recherche opérationnelle, ROADEF’2000, Nantes,
France, 26-28 January 2000.
[107] M. Sevaux et S. Dauzère-Pérès. Genetic algorithms to minimize the weigh-

Research, 151(2) :296–306, 2003.
106
Références
[108] M. Sevaux et Y. Le Quéré. Solving a robust maintenance scheduling pro-

blem at the French railway company. Technical Report LAMIH/SP-2003-3,
University of Valenciennes, CNRS UMR 8530, LAMIH/SP, April 2003. Sub-
mitted.
[109] M. Sevaux et Y. Mineur. Continuous optimisation by GA for a curve fit-
ting problem in autmobile industry. Technical report, University of Valen-
ciennes, CNRS UMR 8530, LAMIH/SP, 2004. Submitted.
[110] M. Sevaux et K. Sörensen. Genetic algorithm for robust schedules. In Pro-
ceedings of 8th International Workshop on Project Management and Scheduling,
PMS’2002, pages 330–333, Valencia, Spain, 3-5 April 2002. ISBN 84-921190-
5-5.
[111] M. Sevaux et K. Sörensen. A genetic algorithm for robust schedules in a
just-in-time environment with ready times and due dates. 4OR – Quaterly
journal of the Belgian, French and Italian Operations Research Societies, 2(2) :
129–147, 2004.
[112] M. Sevaux, K. Sörensen, et Y. Le Quéré. Flexibilité et robustesse en ordonnance-
ment, chapter Métaheuristiques pour la planification et l’ordonnancement
robuste, pages 91–110. Hermes, 2004. To appear.
[113] M. Sevaux et P. Thomin. Efficient heuristic and tabu search for parallel ma-
chine scheduling. In ORBEL Conference, Antwerp, Belgium, 29-30 January
2001.
[114] M. Sevaux et P. Thomin. Heuristics and metaheuristics for parallel ma-
chine scheduling : a computational evaluation. In Proceedings of 4th Meta-
heuristics International Conference, MIC 2001, pages 411–415, Porto, Portugal,
16-20 July 2001.
[115] M. Sevaux et P. Thomin. Heuristics and metaheursitics for a parallel ma-
chine scheduling problem : a computational evaluation. Technical Re-
port LAMIH/SP-2001-2, University of Valenciennes, CNRS UMR 8530, LA-
MIH/SP, November 2001. Submitted.
[116] M. Sevaux et P. Thomin. Parallel machine scheduling : a metaheuristic com-
putational evaluation. Séminaire invité, Institut de Recherche en Commu-
nication et Cybernétique de Nantes, IRCCyN (in French), 19 April 2001.
[117] M. Sevaux et P. Thomin. Parallel machine scheduling : a (meta)heuristic
computational evaluation. Séminaire invité, Groupe MOST (in French), 5
April 2001.
[118] M. Sevaux et P. Thomin. Recherche taboue améliorée pour l’ordon-
nancement sur machines parallèles. In Actes de la 3ième conférence internatio-
nale francophone de MOdélisation et de SIMulation, MOSIM’01, pages 829–833,
Troyes, France, 25-27 April 2001. ISBN 1-56555-212-1.
107
Références
[119] M. Sevaux et P. Thomin. Scatter search and ga : a one machine scheduling

problem comparison. In The sixteenth triennial conference of international fede-
ration of operational research societies, IFORS’2002, Edinburgh, UK, 6-13 July
2002.
[120] K. Sörensen. Distance measures based on the edit distance for permutation-
type representations. In A. Barry, editor, Proceedings of the Workshop on Ana-
lysis and Design of Representations and Operators (ADoRo), GECCO Conference,
pages 15–21, Chicago, 2003.
[121] K. Sörensen. A framework for robust and flexible optimisation using metaheu-
ristics with application in supply chain design. PhD thesis, University of Ant-
werp, 2003.
[122] K. Sörensen et M. Sevaux. GA|PM : genetic algorithms with population

[123] K. Sörensen et M. Sevaux. Robust and flexible vehicle routing in practical
situations. In Proceedings of 5th triennial symposium on transportation analysis,
TRISTAN V, Le Gosier, Guadeloupe, France, 13-18 June 2004. To appear.
[124] E.G. Talbi. A taxonomy of hybrid metaheuristics. Journal of Heuristics, 8 :
541–564, 2002.
[125] P.J.M. Van Laarhoven et E.H.L. Aarts. Simulated Annealing : Theory and Ap-
plications. Kluwer Academic Publishers, Boston, 1987.
[126] R.V. Vidal, editor. Applied simulated annealing, volume 396 of LNEMS.
Springer-Verlag, Berlin, 1993.
[127] C. Voudouris. Guided local search for combinatorial optimization problems. PhD
thesis, Dept. of Computer Science, University of Essex, Colcherster, UK,
1997.
[128] C. Voudouris et E. Tsang. Guided local search. Technical Report TR CSM-
247, University of Essex, UK, 1995.
[129] C. Voudouris et E. Tsang. Guided local search and its application to the
travelling salesman problem. European Journal of Operational Research, 113 :
469–499, 1999.
108
Troisième partie
Sélection de publications
110
Sélection de publications
Description des publications

Pour conclure ce document, nous proposons ici quelques publications. A l’ex-
ception d’une seule [3], toutes les références sont publiées [1, 5] ou à paraı̂tre [2, 4].
Elles sont à notre sens représentatives de notre activité récente.
Concernant les métaheuristiques proprement dites, [1] présente l’étude de
plusieurs opérateurs d’un algorithme génétique. Le problème traité est celui de
l’ordonnancement à une machine. [2] propose un algorithme génétique adapté
pour résoudre le même problème d’ordonnancement dans le cas où les données
sont incertaines. L’utilisation des métaheursitiques pour résoudre un problème
bi-critère de tournées de véhicules sur arcs est présenté dans [4].
Un travail de plus grande envergure présenté dans [3] propose la description
d’une nouvelle métaheuristique qui gère de manière précise la population. Même
si cet article n’est pas encore accepté, il mérite d’être présenté car il se trouve être
une nouvelle source de travail et une nouvelle piste importante de recherche.
Enfin, nous souhaitions présenter un travail déconnecté des métaheuristiques
[5] mais qui nous tient à cœur particulièrement. La résolution par relaxation La-
grangienne d’un problème d’ordonnancement à une machine. Ce article est là
pour rappeler qu’il n’existe pas que les métaheuristiques et qu’il est important de
pouvoir proposer des techniques de résolution alternatives (ou pourquoi pas des
techniques à hybrider avec les métaheuristiques).
1 European Journal of Operational Research (2003)
[1] M. Sevaux and S. Dauzère-Pérès. Genetic algorithms to minimize the weigh-

Research, 151(2) :296–306, 2003.
Abstract : The general one-machine scheduling problem is stron-
gly N P-Hard when the objective is to minimize the weighted
number of late jobs. Few methods exist to solve this problem. In
an other paper, we developed a Lagrangean relaxation algorithm
which gives good results on many instances. However, there is
still room for improvement, and a metaheuristic might lead to bet-
ter results. In this paper, we decided to use a Genetic Algorithm
(GA). Although a GA is somewhat easy to implement, many va-
riations exist, and we tested some of them to design the best GA
for our problem. Three different engines to evaluate the fitness of a
chromosome are considered, together with four types of crossover
operators and three types of mutation operators. An improved GA
is also proposed by applying local search on solutions determined
111
from the chromosome by the engine. Numerical experiments on
different tests of instances are reported. They show that starting
from an initial population already containing a good solution is
very effective.
2 4OR – Quaterly journal of the Belgian, French and Italian

Operations Research Societies (2004 à paraı̂tre)
[2] M. Sevaux and K. Sörensen. A genetic algorithm for robust schedules in a

just-in-time environment with ready times and due dates. 4OR – Quaterly
journal of the Belgian, French and Italian Operations Research Societies, 2004. In
Press.
Abstract : Computing a schedule for a given single machine pro-
blem is often difficult for irregular criteria, but when the data are
uncertain, the problem is much more complicated. In this paper,
we modify a genetic algorithm to compute robust schedules when
release dates are subject to small variations. Two types of robust-
ness are distinguished : quality robustness or robustness in the ob-
jective function space and solution robustness or robustness in the
solution space. The modified genetic algorithm is applied to a just-
in-time scheduling problem, a common problem in several indus-
tries.
3 Journal of Heuristics (2004 soumis)

Abstract : We present a new type of evolutionary algorithms cal-
led genetic algorithms with population management or GA|PM .
The main distinguishing feature of GA|PM is the use of distance
measures to control the diversity of a small population. Combi-
ned with a local search procedure to ensure the high quality of
individual solutions in the population, population management
strategies can be developed to avoid slow or premature conver-
gence and achieve excellent performance. The new algorithm is
tested on two problems : the multidimensional knapsack problem
and the weighted tardiness single machine scheduling problem.
On both problems, population management is shown to be able to
improve the performance of an evolutionary algorithm.
112
4 Computers and Operations Research (2004 à paraı̂tre)
[4] P. Lacomme, C. Prins, and M. Sevaux. A genetic algorithm for a bi-objective

capacitated arc routing problem. Computers and Operations Research, 2004. In
Press.
Abstract : The Capacitated Arc Routing Problem (CARP) is a very
hard vehicle routing problem for which the objective — in its clas-
sical form — is the minimisation of the total cost of the routes.
In addition, one can seek to minimize also the cost of the longest
trip. In this paper, a multi-objective genetic algorithm is presented
for this more realistic CARP. Inspired by the second version of the
Non-dominated Sorted Genetic Algorithm framework, the proce-
dure is improved by using good constructive heuristics to seed
the initial population and by including a local search procedure.
The new framework and its different flavour is appraised on three
sets of classical CARP instances comprising 81 files. Yet designed
for a bi-objective problem, the best versions are competitive with
state-of-the-art metaheuristics for the single objective CARP, both
in terms of solution quality and computational efficiency : indeed,
they retrieve a majority of proven optima and improve two best-
known solutions.
5 Naval Research Logistics (2003)

2003.
Abstract : This paper tackles the general single machine schedu-
ling problem, where jobs have different release and due dates and
the objective is to minimize the weighted number of late jobs. The
notion of master sequence is first introduced, i.e., a sequence that
contains at least an optimal sequence of jobs on time. This mas-
ter sequence is used to derive an original mixed-integer linear pro-
gramming. By relaxing some constraints, it is possible to propose
a Lagrangean relaxation algorithm which gives both a lower and
upper bound. Although the duality gap becomes larger with the
number of jobs, it is possible to solve problems of more than 100
jobs, and some computational results are presented.
113
114
European Journal of Operational Research 151 (2003) 296–306
www.elsevier.com/locate/dsw
Genetic algorithms to minimize the weighted number

of late jobs on a single machine
a,* b
Marc Sevaux , Stephane Dauzere-Peres
a
Universit
e de Valenciennes, Mont Houy, CNRS, UMR 8530, LAMIH/SP, F-59313 Valenciennes Cedex, France
b
IRCCyN, Ecole des Mines de Nantes, CNRS, UMR 6597, La Chantrerie, BP 20722, F-44307 Nantes Cedex 3, France
Abstract
The general one-machine scheduling problem is strongly NP-Hard when the objective is to minimize the weighted
number of late jobs. Few methods exist to solve this problem. In an other paper, we developed a Lagrangean relaxation
algorithm which gives good results on many instances. However, there is still room for improvement, and a meta-
heuristic might lead to better results. In this paper, we decided to use a genetic algorithm (GA). Although a GA is
somewhat easy to implement, many variations exist, and we tested some of them to design the best GA for our problem.
Three different engines to evaluate the fitness of a chromosome are considered, together with four types of crossover
operators and three types of mutation operators. An improved GA is also proposed by applying local search on so-
lutions determined from the chromosome by the engine. Numerical experiments on different tests of instances are
reported. They show that starting from an initial population already containing a good solution is very effective.
Ó 2003 Elsevier B.V. All rights reserved.
Keywords: Scheduling; Single machine; Weighted tardy jobs; GA; Hybrid GA
1. Introduction time), and late otherwise. Finally, a weight wi is

associated to each job Ji , and the objective in this
The general one-machine sequencing problem paper is to minimize the weighted number of late
can be described as follows: a set J of n jobs jobs (or equivalently to maximize the weighted
ðfJ1 ; . . . ; Jn gÞ has to be scheduled on a single ma- number of early jobs).
chine, and each job Ji has a release date ri , a
processing time pi and a due date di . Moreover, the Remark 1. A well-known and important remark is
machine can only process one job at a time, and that there is always an optimal schedule in which
preemption is not allowed. A scheduled job com- late jobs are sequenced after all the early jobs. A
pleted before its due date is said to be early (or on partial sequence of early jobs describes completely
a feasible solution.
*
Corresponding author. This
P single machine scheduling problem, noted
E-mail addresses: marc.sevaux@univ-valenciennes.fr (M. 1jrj j wj Uj in the standard classification [10], is
Sevaux), stephane.dauzere-peres@emn.fr (S. Dauzere-Peres). strongly NP-Hard [14]. When release dates are all
0377-2217/$ - see front matter Ó 2003 Elsevier B.V. All rights reserved.
doi:10.1016/S0377-2217(02)00827-5
M. Sevaux, S. Dauzere-Peres / European Journal of Operational Research 151 (2003) 296–306 297
P
equal to zero ð1k wj Uj Þ, the problem is still Section 2 introduces some general consider-
NP-Hard. When all weights are equal ð1jrj j
P ations about GA. Section 3 is devoted to finding a
Uj Þ, the problem remains NP-Hard, but can be good combination for the engine, and the cross-
solved by an Oðn log P nÞ algorithm if all release over and mutation operators. The genetic local
dates are equal ð1k Uj Þ [15]. If release and due search (GLS) is presented in Section 4 and nu-
dates are similarly ordered ðri < rj ) di 6 dj , merical experiments are reported in Section 5.
8ðJi ; Jj ÞÞ Kise et al. [12] gave an Oðn2 Þ algorithm
and Lawler [13] an Oðn log nÞ algorithm for the
same special case. 2. Genetic algorithms
However, some exact approachesPhave recently
Pproposed for the problem 1jrj j Uj [2,6]. For
been The basic concepts of GAs, introduced by
1k wj Uj , Lawler [13] showed that MooreÕs al- Holland [11], have been successfully applied to
gorithm ([15]) could be applied when process- solve many combinatorial optimization problems.
ing times and weights are agreeable ðpi < pj ) We refer the reader to a recent feature article from
wi P wj , 8ðJi ; Jj ÞÞ. Finally, branch-and-bound pro- Reeves [20] for an extensive description of GAs in
cedures have been developed to solve the P case operations research.
where all release dates are equal ð1k wj Uj Þ
[21,17]. 2.1. General considerations
A new and efficient Lagrangean relaxation al-
gorithm (LRA) has been presented by Peridy et al. A GA tries to mimic the genetic behavior of a
[16] and their lower bound will be used in the species. The biggest difference with other meta-
computational experiments. In a recent work, heuristics (like Tabu search (TS) or simulated an-
Dauzere-Peres and Sevaux [7] developed a nealing (SA)) is that GA maintains a population of
Lagrangean relaxation heuristic to solve the solutions rather than a unique current solution.
problem in the general case. Because the Lagran- Solutions are coded as finite-length strings called
gean relaxation heuristic gives at the same time a chromosomes and a measure of their adaptation
feasible solution (an upper bound) and a lower (the fitness) is computed by an engine. Starting
bound, a measure of the quality of a solution can from an existing population, each iteration gener-
be given for each instance. The same heuristic ates new chromosomes by applying operators (like
can be used to solve the problem when weights are crossover and mutation) to two ‘‘randomly’’ chosen
equal and thus has been compared to optimal so- parents. The main advantage of a GA is its in-
lutions given by a branch-and-bound algorithm [6]. trinsic parallelism, which allows the exploration of
Because we used generic parameters for the LRA a larger solution space than in TS or SA. With an
(the parameters are the same for all the randomly appropriate design, GAs can give very good re-
generated instances), in some test problems, the sults.
gap between the solution and the lower bound is
rather large and can be reduced. 2.2. Chromosome encoding
This paper studies different classical genetic al-
gorithm (GA) techniques, combines them and In [18], Portmann describes how GAs can be
presents the first hybrid genetic algorithm (HGA) applied to scheduling problems, and some very
that Psolves efficiently the NP-Hard problem interesting hints for building a specific GA can be
1jrj j wj Uj . An initial solution given by a LRA is found.
used as a good starting point but, when the size of Often, when a GA is implemented to solve a
the instances increases, the HGA solution almost one-machine scheduling problem, the representa-
always improves the best solution found so far. tion of a chromosome is a permutation of the jobs
This suggests that the HGA is very efficient for ‘‘J1 ; J2 ; . . . ; Jn ’’. This representation is easy to
large size instances. handle and has been adopted for our study.
298 M. Sevaux, S. Dauzere-Peres / European Journal of Operational Research 151 (2003) 296–306
2.3. Population management offspring. Afterwards, the population is sorted

again according to the fitness values.
Starting from an initial population of size m,
two ways of managing the population are usually
considered: either a new population of size m is 3. Designing a genetic algorithm
generated and replaces the previous population, or
only one offspring is generated and replaces one 3.1. Fitness
chromosome in the existing population. The for-
mer method is named population replacement and For a given chromosome (i.e., a permutation of
the latter incremental replacement. the n jobs), a feasible solution, corresponding to
In population replacement, it may happen that the sum of the weights of the late jobs, has to be
the best chromosome in the initial population, i.e., computed, i.e., a feasible schedule has to be de-
the best current solution for our problem, is no termined. This is called the engine. Three engines,
longer in the new population, and thus will not be denoted by ‘‘list’’, ‘‘non-delay (ND)’’ and ‘‘Fifo’’
used again by the crossover operator. This is the have been tested. Since a chromosome is a per-
reason why most researchers prefer incremental mutation of n jobs, an engine is nothing else than a
replacement, also called elitism or population list algorithm.
overlaps [8]. In the sequel, only incremental re- The list engine is very simple. Jobs are taken in
placement will be considered. the order of the chromosome and sequenced con-
secutively when on time. A job which is late if
2.4. Reproduction mechanism sequenced at the current position is noted late, and
will be sequenced after the jobs on time (see Re-
At every iteration, the selection of the two mark 1). Note that, if a job has been sequenced at
parents is done according to a simple ranking time t and has left an idle period before t, the list
mechanism [19] which gives more chances to engine will never schedule a job in this idle period.
chromosomes with high fitness values to be used This engine builds semi-active schedules, i.e., in
for reproduction. The population is sorted in as- which jobs are sequenced as soon as possible
cending order of fitness values (ascending order of without changing the order of the jobs on the
the objective function), so the best individual is machine. Baker [1] named this engine the semi-
ranked first. The selection of the first parent is active schedule generator. The evaluation function
made using the probability distribution 2ðm þ runs in OðnÞ.
1 kÞ=ðmðm þ 1ÞÞ where k is the kth chromosome The ND engine builds non-delay schedules as
in the ascending order of fitness values, and m the the BakerÕs non-delay schedule generator [1]. The
size of the population. Using this technique, the idea is to sequence a job each time it is possible
median chromosome ðm þ 1Þ=2 has the probability with no delay. When more than one job can be
1=m of being selected, while the best individual has sequenced at the same time, the order given by the
the probability 2=ðm þ 1Þ (roughly twice the me- chromosome determines the jobs that will be se-
dian). The second parent is randomly selected with quenced first. Jobs that cannot be scheduled on
an uniform distribution. time are sequenced after all the early jobs. This
A crossover operator is applied on the two engine might exclude the optimal solution. This
parents and a new offspring is generated. Nor- engine is actually a modification of the Giffler al-
mally, two offspring are created but only one is gorithm, originally designed for problems without
kept in the algorithm. To keep the size of the release dates and regular criteria (see [9]). The
population constant, the new offspring should re- complexity of the ND engine is Oðn2 Þ.
place another individual in the existing population. The Fifo engine builds active schedules and is
In our procedure, an individual is randomly equivalent to the strict order generator proposed
chosen after the median and replaced by the new first by Carlier [4]. A list L of idle periods where
jobs can be scheduled is maintained. At the be- The PMX operator (partial matching crossover)
ginning, the list L is the interval ½0; þ1½. Each job creates a matching between the two parents. In the
in the chromosome is sequenced at the earliest example, job J7 is replaced by job J3 or the opposite,
time, and the list L of idle periods is updated. Jobs J4 by J6 and J5 by J2 . Parent 2 is scanned from the
that can no longer be sequenced are late and de- index 0 and, whenever possible, a job is replaced by
layed after the jobs on time. Once a job is se- its matching. In the offspring, the first job of parent
quenced on time, its starting time is fixed and the 2, J4 , is replaced by J6 , job J3 by J7 , J7 by J3 , etc.
job cannot be moved in the remainder of the al- The X1 operator (one-point crossover) is a
gorithm, i.e., push and pull operations are not al- particular case of the LOX operator where the first
lowed with this engine. The complexity of the crossover point is the index 0.
engine is Oðn2 Þ.
3.3. Mutation
3.2. Crossover
The selection process combined with the cross-
To determine the best GA for our problem, over operators help the population to converge.
four types of crossover operators have been tested. However, after many iterations in the algorithm,
Those operators are commonly used in GAs and the population could be composed of local optima
particularly in scheduling problems: ‘‘OX’’, and needs some diversity to reach new points in
‘‘LOX’’, ‘‘PMX’’ (two-point crossover operators) the solution space. This is the role of the mutation
and ‘‘X1’’ (or ‘‘1X’’) (one-point crossover opera- operators. In this paper, three types of mutation
tor) (see [18]). operators are evaluated. The ‘‘swap’’ operator
In the following, two parent chromosomes with permutes two contiguous elements. The ‘‘big
seven jobs are used, and Table 1 illustrates the swap’’ operator permutes two randomly chosen
outcome of each two-point crossover operator. elements of the chromosome. The ‘‘shift’’ operator
For the two-point crossover operators, the first takes an element and inserts it after a partial se-
point is between positions 2 and 3 and the second quence which is then left-shifted.
point between positions 5 and 6. These operators
create two offspring but only one is kept. 3.4. Other parameters
For OX, LOX and PMX crossover operators,
two crossover points are randomly chosen. In the For our study, the number of individuals in the
OX and LOX operators, the partial sequence of population is set to m ¼ 5 þ n=5, and the mutation
the first parent between these two points are cop- rate to 0.25 (i.e., on average, only 1 out of 4
ied into the new offspring. In the OX operator, the chromosomes is mutated). Two chromosomes are
second parent is scanned from the index of the identical if the permutation of the n jobs is the
second crossover point and filled up with the jobs same. Two completely different chromosomes can
that are not already in the offspring, starting after have two identical fitness values. This is why we
the second crossover point. The LOX operator fills decided to allow a maximum of three clones (three
up the offspring in the same manner but starts in identical fitness values in the population).
the offspring from index 0. Due to Remark 1, in any solution, late jobs are
always scheduled after the jobs on time. It is pos-
sible to keep that ‘‘order’’ by reordering the
Table 1 chromosome after the evaluation. Early jobs are
Results of the two-point crossover operators
scheduled first and late jobs after in the order of
Parent 1 1 2 3 4 5 6 7 the evaluated chromosome. This feature (called
Parent 2 4 3 7 6 2 5 1
reorder) will be taken into account for the combi-
Offspring OX 6 2 3 4 5 1 7 nation of parameters.
Offspring LOX 7 6 3 4 5 2 1 For the stopping conditions, the maximum
Offspring PMX 6 7 3 4 5 2 1 number of iterations grows with the number of
jobs (size of the instance) ItMax ¼ 50 n m, beginning of the chromosome i.e., a partial se-
and the maximum number of iterations without quence of early jobs. This is one of the reasons why
improvement is ItMax/10. this combination works well.
The big swap mutation operator could be re-
3.5. Choosing a combination of parameters placed by the shift mutation operator since the
results are quite equivalent. The swap mutation
To run one version of the GA, an engine, a operator does not perturbate enough the chro-
crossover operator and a mutation operator have mosome to escape from local optima.
to be chosen. Optionally, we can reorder a chro-
mosome after its evaluation according to two
partial sequences of early and late jobs. Three
engines, four crossover operators and three mu- 4. Genetic local search
tation operators and the reorder option are avail-
able. Combining these parameters leads to 72 GAs are general search techniques. Sometimes
differents possible test runs. a chromosome may lead to a solution very close to
To choose the best combination of parameters, an optimal solution but, due to the myopic search
we have generated a set of 1000, 100-job problems of the GA, will not find this optimal solution. An
whose optimal solutions were known. The devia- hybrid genetic algorithm (HGA) would help in this
tion from the optimal solution has been measured case, i.e., a technique that combines at the same
for each instance and used to determine the best time a GA and another improvement method in-
operators and parameters. side the GA.
The best combination is obtained when the re-
ordering of a chromosome occurs at the end of its 4.1. Local improvement
evaluation, associated with the X1 crossover op-
erator that keeps a partial sequence of early jobs. The GLS chosen here, is a simple method that
The best engine is the Fifo engine, and the muta- improves each chromosome of the GA. The
tion operator big swap gives the best results. structure of the GLS developed for our study is the
The ‘‘ND’’ engine builds non-delay schedules, following:
and for an irregular criteria it can exclude the
optimal solution. Nevertheless, on average, this Step 1 Initialization: create an initial population
engine gives good results. Since our objective is to of size m.
reach the optimal solution and not give a good Step 2 Improvement (1): apply a local search to
result, we will not keep it for the final version of improve each chromosome of the initial popula-
the GA. The list engine is too simple and cannot tion (each chromosome is now a local opti-
improve the sequences of early jobs even if there is mum).
a large idle period at the beginning. The Fifo en- Step 3 Recombination: using genetic operators
gine fills up these idle periods when it is possible, (crossover, mutation) create an offspring.
but can be improved again. Step 4 Improvement (2): apply a local search to
Of course, these crossover operators could be improve the offspring.
used in a GA with different probabilities. For ex- Step 5 Replacement: replace a randomly chosen
ample, the ND engine or the Fifo engine could be individual in the population by the new off-
chosen with probability 1/2, improving on average spring.
the solution (ND engine) and not excluding the Step 6 Stop: if stopping conditions are not met
optimal solution (Fifo engine). This is a topic for go to Step 3.
future investigations.
When the reorder parameter is active, the early Performing the local search in Steps 2 and 4
jobs are always sequenced first. The X1 crossover requires the definition of a neighborhood struc-
operator keeps a partial sequence starting from the ture, and a strategy to search this neighborhood.
4.2. Neighborhood stances has been used in the Lagrangean relaxation

algorithm proposed in [7]. The second one comes
Since the objective is to minimize the weighted from [3]. On this last set of instances, another very
number of late jobs, each chromosome is com- efficient lower bound from Peridy et al. [16] is used
posed of a list of early jobs and a list of tardy jobs to measure the quality of our results.
after the execution of a chosen engine. A neighbor
of a current solution is obtained by exchanging a 5.1. Instance generator
job in the list of tardy jobs with a job in the list of
early jobs. The first set of instances has been used in [5–7].
It has been extended for the weighted case as fol-
4.3. Strategy lows. For each job Ji , a processing time pi is ran-
domly generated in the interval [1, 100] and a
Many local search procedures can be used in weight wi is generated in the interval [1, 10] or in
Steps 2 and 4 of the GLS. Some authors tried Tabu [1, 99]. As in [5], two parameters K1 and K2 are
Search or SA and obtained good results. The used, and taken in the set {1, 5, 10, 20}. Because
problem is that we do not want to spend too much we want data to depend on the number of jobs n,
CPU time in completing the local search, in order to the release date ri is randomly generated in the
ensure that the GA can still perform each iteration interval ½0; K1 n, and the due date in the interval
quickly enough. This is why we chose a descent. ½ri þ pi ; ri þ pi þ K2 n. The algorithm was tested for
To improve the current solution, a greedy al- n 2 f20; 40; 60; 80; 100; 120; 140g. For each com-
gorithm is used. A move is made if and only if it bination of n, K1 , and K2 , 10 instances are gener-
strictly improves the current value of the objective ated, i.e., for each value of n, 160 instances are
function. The algorithm to reach a local optimum generated with a weight randomly choosen in
is described below. [1, 10] and 160 instances with a weight in [1, 99]. In
Repeat the following steps until every late job [7], weights are only generated in the interval
has been examined: [1, 10].
1. Choose a job Jt in the ‘‘initial tardy list’’. 5.2. Numerical results of GA and GLS
2. Start from the first job Je in the ‘‘early list’’.
3. If Jt can replace Je and the total weight of the In this section, the results of the Lagrangean
tardy jobs decreases, set Jt on time and add Je relaxation algorithm (named LRA in the sequel)
to the ‘‘final tardy list’’ (Je is definitively late are recalled. The genetic algorithm and the genetic
and will not be chosen again in 1), go to 1. local search are denoted by GA and GLS, re-
4. Choose the next job Je in the early list and go to spectively. To evaluate the quality of a solution,
3. If no such job exists, add Jt to the final tardy the usual relative gap (rg) between an upper bound
list. (UB) and a lower bound (LB) is computed using
the following formula:
The complexity depends on the number of jobs
in the ‘‘tardy list’’. If t denotes this number, the UB LB
rg ¼ :
complexity is Oðtðn tÞÞ and the maximum is UB
reached for t ¼ n=2. The overall complexity of the
local search procedure is Oðn2 Þ. 5.2.1. Results of the Lagrangean relaxation algo-
rithm
Table 2 gives the relative gap for the upper
5. Computational results bound obtained with the LRA. The first column
gives the number of jobs. The next three columns
In this section, two sets of instances are used to give the average value, the standard deviation and
evaluate our metaheuristic. The first set of in- the maximum value of rg. The last two columns
Table 2 give the CPU time in seconds before the LRA

Relative gap for the LRA stops (average and maximum CPU time). It can be
No. of Relative gap (rg) CPU time (seconds) observed that the relative gap increases with the
jobs Mean StDev Max Mean Max number of jobs. For n ¼ 140, the relative gap is
20 0.10 0.01 0.49 0.82 2.52 around 33%. This is one of the reasons that mo-
40 0.09 0.01 0.64 2.98 8.31 tivated the development of a metaheuristic.
60 0.13 0.01 0.61 6.86 18.45
80 0.17 0.04 0.71 12.07 32.12
5.2.2. Results of the genetic algorithm
100 0.22 0.06 0.85 19.69 49.05
120 0.27 0.08 1.00 28.94 74.35 The GA computes an upper bound for which
140 0.33 0.09 0.92 39.97 98.40 the relative gap is on average bigger than the re-
sults of the LRA (see Table 3). The CPU times are
smaller but not far from those of the LRA. Thus,
Table 3
the GA used alone is not more efficient than the
Relative gap for the GA LRA.
No. of Relative gap (rg) CPU time (seconds)
jobs 5.2.3. Results of the genetic local search
Mean StDev Max Mean Max
Because it requires more computation, the time
20 0.18 0.01 0.79 0.09 1.09
to run the algorithm is bigger for the GLS than for
40 0.19 0.01 0.86 0.36 1.42
60 0.21 0.02 0.73 1.14 3.21 the simple GA. But, for instances with more that
80 0.24 0.03 0.73 2.90 8.17 100 jobs, Table 4 shows that the solution is better
100 0.28 0.04 0.84 6.42 16.11 than the upper bound given by the LRA. How-
120 0.31 0.06 1.00 12.38 44.28 ever, even if the solution is better, a relative gap of
140 0.36 0.07 0.89 21.29 65.00
21% is still large and can probably be improved
again.
Table 4
5.3. Impact of a good initial solution
Relative gap for the GLS
No. of Relative gap (rg) CPU time (seconds)
It is known that a good initial solution for a
jobs genetic-based algorithm usually leads to better
Mean StDev Max Mean Max
results. The upper bound of the LRA is used as a
20 0.14 0.01 0.63 0.11 1.17
chromosome in the initial population of the GA
40 0.15 0.01 0.78 0.88 3.10
60 0.16 0.02 0.61 3.92 13.48 and the GLS. These two algorithms will be de-
80 0.19 0.03 0.71 12.07 51.34 noted by GALag and GLSLag, respectively.
100 0.21 0.04 0.77 73.15 280.80 Table 5 gives the results for GALag and com-
120 0.26 0.06 1.00 63.48 238.82 pares them to the simple LRA. The first column is
140 0.30 0.07 0.87 122.56 528.25
the number of jobs. The next three column give the
Table 5
Improvement of the genetic algorithm (GALag)
No. of jobs GALag relative gap Improvement CPU time (seconds)
Mean StDev Max (%) Mean (%) Max (%) Mean Max
20 0.09 0.01 0.49 22.5 4.2 16.6 0.92 2.28
40 0.08 0.00 0.56 28.1 3.1 19.6 3.86 8.09
60 0.11 0.01 0.55 48.8 3.7 22.5 7.23 19.03
80 0.14 0.02 0.64 64.4 6.7 43.4 12.12 36.14
100 0.18 0.04 0.79 72.2 9.1 42.7 22.00 54.11
120 0.23 0.07 1.00 80.9 11.1 61.6 33.75 76.98
140 0.28 0.08 0.88 90.0 12.4 54.1 49.04 116.35
Table 6
Improvement of the genetic local search (GLSLag)
No. of jobs GLSLag relative gap Improvement CPU time (seconds)
Mean StDev Max (%) Mean (%) Max (%) Mean Max
20 0.09 0.01 0.49 18.8 4.5 16.6 0.86 2.57
40 0.08 0.00 0.64 24.4 2.7 10.7 3.26 8.61
60 0.11 0.01 0.54 46.3 3.8 20.2 8.21 20.80
80 0.14 0.02 0.64 63.8 6.9 39.3 16.30 50.45
100 0.18 0.04 0.78 74.4 9.2 42.6 30.12 80.65
120 0.23 0.07 1.00 82.5 11.3 54.2 51.45 174.88
140 0.27 0.08 0.87 90.6 13.4 52.9 87.18 237.69
Table 7
Comparison of the algorithms ðn ¼ 10; . . . ; 50Þ
Algorithm LRA GA GLS GALag GLSLag
n ¼ 10
Dev. from LBLag (%) 10.3 29.4 32.6 10.3 10.3
Dev. from LBPer (%) 1.7 22.0 24.9 1.6 1.6
Optimum found (%) 91.1 52.2 53.3 93.3 93.3
# of time LRA is strictly improved (%) 2.2 2.2
Average CPU time (seconds) 0.32 0.02 0.02 0.33 0.33
Maximum CPU time (seconds) 0.61 0.09 0.05 0.63 0.63
n ¼ 20
Dev. from LBLag (%) 12.2 27.5 25.0 11.8 11.6
Dev. from LBPer (%) 4.2 20.7 18.0 3.7 3.6
Optimum found (%) 61.1 11.1 17.8 66.7 65.6
n ¼ 30
Dev. from LBLag (%) 12.8 26.6 22.4 12.3 12.4
Dev. from LBPer (%) 3.2 18.5 13.8 2.6 2.7
Optimum found (%) 46.7 5.6 12.2 54.4 54.4
n ¼ 40
Dev. from LBLag (%) 9.4 23.7 20.9 9.2 9.2
Dev. from LBPer (%) 2.1 17.3 13.4 1.8 1.8
Optimum found (%) 41.1 5.6 10.0 46.7 44.4
n ¼ 50
Dev. from LBLag (%) 11.1 25.6 19.5 10.7 10.6
Dev. from LBPer (%) 3.2 18.6 12.1 2.7 2.6
Optimum found (%) 35.6 2.2 4.4 40.0 37.8
relative gap (as in Tables 3 and 4). The three col- havior shows that if GLSLag is stopped too early,
umns ‘‘improvement’’ give the number of time (in GALag may give better results because more it-
%) GALag improves the initial solution of LRA, erations can be done in the same amount of time.
the average and maximum value of the improve- However, even when small CPU times are allowed,
ment ððUBLRA UBGALag Þ=UBLRA Þ. The last two GLSLag almost always dominates GALag. For a
columns give the average and maximum CPU time specific 100-job instance, GALag performs 14853
in seconds (time for the LRA þ GA). iterations in 11.59 seconds and the best solution is
Table 6 gives the same results for GLSLag and 2143. GLSLag performs only 3555 iterations in
compares them to the ones of LRA. GLSLag gives 18.14 seconds and the best solution is 2091. GA-
the best results on these tests (the solution is the Lag obtains a better solution than GLSLag before
best almost 90% of the time) except for the CPU a time limit of 2.56 seconds, but is surpassed by
time (however still reasonable). The general be- GLSLag afterwards. This is not as obvious in the
Table 8
Comparison of algorithms ðn ¼ 60; . . . ; 100Þ
Algorithm LRA GA GLS GALag GLSLag
n ¼ 60
Dev. from LBLag (%) 10.8 23.4 21.7 10.2 10.2
Dev. from LBPer (%) 3.7 17.0 12.2 2.9 2.9
Optimum found (%) 24.4 4.4 7.8 32.2 36.7
n ¼ 70
Dev. from LBLag (%) 10.6 23.7 20.9 9.7 9.8
Dev. from LBPer (%) 3.8 17.6 12.3 2.9 3.0
Optimum found (%) 23.3 2.2 4.4 32.2 31.1
n ¼ 80
Dev. from LBLag (%) 10.3 23.6 22.4 9.6 9.6
Dev. from LBPer (%) 3.8 17.6 12.2 2.7 2.8
Optimum found (%) 25.6 1.1 2.2 26.7 28.9
n ¼ 90
Dev. from LBLag (%) 10.2 23.8 22.3 9.3 9.4
Dev. from LBPer (%) 3.5 17.8 12.0 2.6 2.6
Optimum found (%) 23.3 1.1 4.4 27.8 28.9
n ¼ 100
Dev. from LBLag (%) 10.9 23.9 18.0 10.0 10.1
Dev. from LBPer (%) 4.1 17.7 11.4 3.0 3.1
Optimum found (%) 20.0 1.1 2.2 24.4 23.3
results of the following section (see Tables 7 and 8) can be determined, i.e., if our solutions are close to
on a different set of instances. optimality or not. This question could be partially
Although GLSLag gives the best results, the answered by using the lower bound proposed in
relative gap for n ¼ 100 is still large (18%). For [16], although duality gaps are still too large to
n ¼ 140, the gap is above 25%. This is probably conclude on the second set of instances.
due to the fact that the lower bounds obtained by Hence, our future research aims at designing an
Lagrangean relaxation are not good enough. Re- exact approach for this problem, which could be
call that general parameters are used for LRA for used to better evaluate the effectiveness of our
every instance. metaheuristic.
5.4. Comparison with a better lower bound

Acknowledgements
Peridy et al. [16] developed a very efficient lower
bound (denoted LBPer) also using Lagrangean The authors would like to thank C. Prins for the
relaxation, but very differently than in [7]. The very useful hints and private conversations that
authors sent us the results on a set of instances motivated this paper. The authors are also grateful
generated in a very different way than our set (see to A. Hertz and M. Laguna for their enlighten-
Baptiste et al. [3]), the main difference being that ment, to P. Thomin for his help in the implemen-
jobs are more constrained (smaller time windows) tation of part of the code, and to two anonymous
in these instances. We ran the different versions of referees for their helpful comments.
the genetic-based algorithms and compared our
results to the lower bound.
The size of these test problems ranges from
n ¼ 10 to 100. Table 7 gives the results for in- References
stances with sizes varying from n ¼ 10 to 50, and
[1] K.R. Baker, Introduction to Sequencing and Scheduling,
Table 8 from n ¼ 60 to 100. They show that the John Wiley, New York, USA, 1974.
duality gap is much smaller with a better lower [2] P. Baptiste, C. Le Pape, L. Peridy, Global constraints for
bound, and that our solutions are thus closer to partial CSPs: A case study of resource and due-date
optimality than what our set of instances indi- constraints, 4th International Conference on Principles and
cated. However, the average deviation from the Practices of Constraint Programming, Pisa, Italy, 1998.
[3] P. Baptiste, A. Jouglet, C. Le Pape, W. Nuijten, A
lower bound is close to 3% for problems of more constraint-based approach to minimize the weighted num-
than 50 jobs, which is still too large to conclude ber of late jobs on parallel machines. Research report 2000/
definitely on the quality of our solution proce- 228, UMR, CNRS 6599, Heudiasyc, France, 2000.
dures. [4] J. Carlier, Problemes dÕordonnancement a contraintes de
resources: Algorithmes et complexites, These dÕEtat, Uni-
versite de Paris 6, 28 mai 1984.
[5] S. Dauzere-Peres, Minimizing late jobs in the general one
6. Conclusion machine scheduling problem, European Journal of Oper-
ational Research 81 (1995) 134–142.
This paper presents a first metaheuristic for a [6] S. Dauzere-Peres, M. Sevaux, An exact method to mini-
NP-Hard one-machine scheduling problem. A mize the number of tardy jobs in single machine schedul-
ing, Research report 99/6/AUTO, Ecole des Mines de
GA has been designed testing various engines and Nantes, France, 1999.
several crossover operators. Consistent improve- [7] S. Dauzere-Peres, M. Sevaux, Using Lagrangean relax-
ments have been obtained on several tests of in- ation to minimize the (weighted) number of late jobs on a
stances, although very good solutions were already single machine, Research report 99/8/AUTO, Ecole des
Mines de Nantes, France, 1999.
provided by a LRA. However, on our test of in-
[8] L. Davis (Ed.), Handbook of Genetic Algorithms, Van
stances, we do not know whether large gaps be- Nostrand Reinhold, New York, 1991.
tween upper and lower bounds are mostly due to [9] S. French, Sequencing and Scheduling: An Introduction to
the duality gap, or to the fact that better solutions the Mathematics of the Job–Shop, Ellis-Horwood, 1990.
[10] R.L. Graham, E.L. Lawler, J.K. Lenstra, A.H.G. Rinnooy [16] L. Peridy, E. Pinson, D. Rivreau, Total cost single-machine
Kan, Optimization and approximation in deterministic scheduling problem with arbitrary release dates, Proceed-
sequencing and scheduling: A survey, Annals of Discrete ings of the Workshop on Project Management and
Mathematics 5 (1979) 287–326. Scheduling, Osnabr€ uck, Germany, April 17–19, 2000, pp.
[11] J.H. Holland, Adaptation in Natural and Artificial Sys- 233–235.
tems, The University of Michigan Press, Ann Arbor, MI, [17] C.N. Potts, L.N. Van Wassenhove, Algorithms for sched-
1975. uling a single machine to minimize the weighted number of
[12] H. Kise, T. Ibaraki, H. Mine, A solvable case of the one- late jobs, Management Science 34 (7) (1988) 834–858.
machine scheduling problem with ready and due times, [18] M.C. Portmann, Genetic algorithms and scheduling: A
Operations Research 26 (1) (1978) 121–126. state of the art and some propositions, Proceedings of the
[13] E.L. Lawler, Knapsack-like scheduling problems, the Workshop on Production Planning and Control, Mons,
Moore–Hodgson algorithm and the Ôtower of setsÕ prop- Belgium, September 9–11, 1996, I–XIV.
erty, Mathematical Computer Modelling 20 (2) (1994) 91– [19] C.R. Reeves, A genetic algorithm for flowshop sequencing,
106. Computers and Operations Research 22 (1) (1995) 5–13.
[14] J.K. Lenstra, A.H.G. Rinnooy Kan, P. Brucker, Com- [20] C.R. Reeves, Genetic algorithms for the operations
plexity of machine scheduling problems, Annals of Discrete researcher, INFORMS Journal on Computing 9 (3)
Mathematics 1 (1977) 343–362. (1997) 231–250.
[15] J.M. Moore, A n job, one machine sequencing algorithm [21] F.J. Villarreal, R.L. Bulfin, Scheduling a single machine to
for minimizing the number of late jobs, Management minimize the weighted number of tardy jobs, IIE Trans-
Science 15 (1) (1968) 102–109. actions 15 (1983) 337–343.
4OR (2004) Regular Paper
DOI: 10.1007/s10288-003-0028-0
A genetic algorithm for robust schedules

in a one-machine environment with ready times
and due dates
Marc Sevaux1 and Kenneth Sörensen2
1 Université de Valenciennes, Le Mont Houy , 59313 Valenciennes, France
(e-mail: marc.sevaux@univ-valenciennes.fr)
2 Universiteit Antwerpen, Middelheimlaan 1, 2020 Antwerp, Belgium
(e-mail: kenneth.sorensen@ua.ac.be)
Received: 19 December 2002 / Accepted: 2 July 2003
Abstract. Computing a schedule for a single machine problem is often difficult, but
when the data are uncertain, the problem is much more complicated. In this paper,
we modify a genetic algorithm to compute robust schedules when release dates
are subject to small variations. Two types of robustness are distinguished: quality
robustness or robustness in the objective function space and solution robustness or
robustness in the solution space. We show that the modified genetic algorithm can
find solutions that are robust with respect to both types of robustness. Moreover, the
risk associated with a specific solution can be easily evaluated. The modified genetic
algorithm is applied to a just-in-time scheduling problem, a common problem in
many industries.
Key words: Quality robustness, solution robustness, single machine scheduling,

weighted number of late jobs, genetic algorithm
Mathematics Subject Classification: 90B99
1 Introduction
Scheduling problems are generally computationally difficult problems and the prob-
lem described in this paper is no exception. The problem becomes even more dif-
ficult when some of the data are stochastic (Daveand Beck 2000; Herroelen and
Correspondence to: K. Sörensen
Quarterly Journal of the Belgian, French

4OR and Italian Operations Research Societies
© Springer-Verlag 2004
2 M Sevaux and K. Sörensen
Leus 2002). Genetic algorithms – and metaheuristics in general – have been found
to provide good approximations of the optimal solution with a relatively small
computational effort (Portmann 1996). Although the literature on applications of
metaheuristics (and genetic algorithms in particular) to scheduling problems is very
large, the number of applications to problems with a stochastic component is rather
small, see, e.g., (Kouvelis and Yu 1997).
One way to deal with stochastic problem data is to find solutions that are robust.
We distinguish two types of robustness. Quality robustness is a property of a solution
whose quality, measured by the objective function value, does not deviate much
from optimality when small changes in the problem data occur.
The second type of robustness is solution robustness and can be described as
robustness in the solution space. When changes in the problem data occur, the
decision maker might be forced to re-optimise the problem. In this case, the quality
of the solution is guaranteed by the optimisation procedure. In some situations
however a solution is preferred that is “close” (in the solution space, not the objective
function space) to the solution currently used. Many manufacturers operate with
a production schedule that repeats itself on a regular basis (e.g., daily or weekly).
When, e.g., a new job needs to be scheduled, the problem is re-optimised, but the
new production schedule should be as similar as possible to the one currently used.
This type of robustness stresses the importance of solution stability. The two types
of robustness are not entirely equivalent in the sense that quality robustness is a
property of a solution that is insensitive to changes in the problem data before these
changes occur, whereas solution robustness refers to the stability of a solution after
changes have occurred.
The object of the single machine scheduling problem considered in this paper
is to determine the order in which a given set of jobs should be executed on a
machine. Each job has a certain release date and cannot be scheduled before this
date. Also, each job has a certain processing time and a due date, before which it
should be finished. Jobs that are not finished before their due date are called late.
The objective of the problem considered in this paper is to minimise the number
of late jobs. A genetic algorithm for this problem was developed by Sevaux and
Dauzère-Pérès (2002).
When a problem has stochastic parameters, the objective function value itself
becomes stochastic. The objective of a stochastic problem has several dimensions.
Some decision makers prefer a solution that has a high quality on average, others
might prefer one that hedges against the worst possible situation. However, most
decision makers like to minimise the risk of ending up with a bad solution, without
necessarily choosing an extremely conservative solution. In this paper, we modify
the GA proposed by Sevaux and Dauzère-Pérès (2002) to be able to deal with
stochastic release dates and show that this GA is able to find solutions that are
both quality robust and solution robust. The proposed approach can be labelled a
sampling method, in that it uses sampling to estimate the robustness of a solution.
We show that this approach is very flexible in that – although our GA is only used for
A genetic algorithm for robust schedules 3
problems with stochastic release dates – other types of stochastic problem data can
be easily incorporated. Simultaneously, the risk preference of the decision maker
can be modelled and a solution can be found that minimises the estimated number
of late jobs and the risk that this solution entails. Solution robustness is obtained
by measuring the distance (in the solution space) to a given base-line solution.
This distance measure is based on the edit distance. Minimising the distance to
a baseline solution and simultaneously maximising the quality of a solution is
obviously difficult. We show in this paper that both objectives can be achieved
simultaneously by letting the GA search for high-quality solutions and measuring
the distance of each solution encountered to the baseline solution. In a second stage,
a solution can be chosen that provides a good trade-off.
This paper is structured as follows.After the literature review, the single machine
scheduling problem is formulated. Section 4 introduces a genetic algorithm for the
deterministic single machine scheduling problem. In Sect. 5, this GA is modified
for the computation of schedules that are both quality robust and solution robust.
Section 6 describes how the modified genetic algorithm can be used to find robust
production sequences in a just-in-time environment. Section 7 reports on some
computational results.
2 Literature review
Single machine scheduling problems have been successfully tackled using ge-
netic algorithms. Genetic algorithms – pioneered by Holland (1975) and Goldberg
(1989), as well as many others—use a Darwinian metaphor to find good solutions
for difficult optimisation problems. A review of genetic algorithms can be found in
Reeves (1993, 1997). For a review of applications of genetic algorithms to schedul-
ing problems we refer to Portmann (1996).
This paper describes a simple GA, developed by Sevaux and Dauzère-Pérès
(2002), that finds good solutions to the deterministic scheduling problem described
above.
Although the literature on deterministic scheduling is several orders of mag-
nitude larger, there is a significant body of work on scheduling under uncertainty.
Two very recent literature surveys are Herroelen and Leus (2002) and Davenport
and Beck (2000). Herroelen and Leus (2002) divide research on scheduling with
uncertainty into six categories. In reactive scheduling, procedures are developed
to re-optimise the schedule when disruptive events occur. Stochastic scheduling is
the application of stochastic programming with recourse to scheduling problems.
GERT network scheduling is used for problems with stochastic evolution structure
and feedback. Scheduling under fuzziness recommends the use of fuzzy numbers
for, e.g., project duration instead of stochastic variables. Proactive (robust) schedul-
ing attempts to develop a schedule, the quality of which is relatively insensitive
to a changing environment. Herroelen and Leus (2002) also discuss some recent
advances in sensitivity analysis for project scheduling. The method developed in
this paper can be categorised under the heading robust scheduling. Davenport and
Beck (2000) distinguish between proactive and reactive approaches. Proactive ap-
proaches include the insertion of redundancy (typically extra buffer time) into the
schedule, the use of probabilistic techniques to determine robust schedules and the
creation of multiple schedules to deal with contingencies.
Although genetic algorithms and meta-heuristics in general have been often and
successfully applied to tackle deterministic scheduling problems, their application
to scheduling under uncertainty is very limited. In this paper, we argue that meta-
heuristics can very easily be adapted to the requirements of a stochastic problem
formulation. Using meta-heuristics for this type of optimisation problem has several
advantages that will be discussed later. That there is a need for robust meta-heuristic
optimisation is recognised in the influential book “Robust discrete optimisation”
(Kouvelis and Yu 1997), when the authors say on p. 354: “We believe that consider-
able more effort should be spent in systematic development of [ . . . ] metaheuristic
frameworks, which with minimal adjustment effort can be applied to a large class of
robust optimisation problems [ . . . ]”. Previous work on robust genetic algorithms
can be found in Tsutsui et al. (1996); Tsutsui and Ghosh (1997); Tsutsui and Jain
(1998); Tsutsui (1999); Branke (1998, 2001), but the applications of this work are
limited to the robust optimisation of continuous mathematical functions.
A genetic algorithm for robust job shop schedules is developed by Leon et al.
(1994b). This algorithm uses an objective function that includes some measure of
robustness. In their problem, the authors assume that machine operations can be
disrupted and that the disrupted operations are restarted immediately after the end
of the disruption period (a right-shift reactive policy). Robustness is defined as a
function of the schedule delay, i.e., the difference between the actual makespan and
the projected makespan. A number of different measures, specifically developed to
estimate this measure of robustness is developed. Leon et al. (1993, 1994b) describe
an approach in which recovering from disruptions is modelled as a game played
against nature.
Hart and Ross (1999a, b) have recently developed artificial immune systems as
a tool for scheduling under uncertainty. The author’s goal is to develop off line a
number of partial schedules that can be used as building blocks to quickly compose
a good schedule when unexpected events occur. The underlying assumption is
that such events (such as machine breakdowns) and the desired reaction of the
scheduling system, are to some extend predictable. The partial schedules found by
this GA therefore constitute a specific piece of domain knowledge, that can be used
when disruptions occur.
Jensen (2001) introduces several novel ideas. In neighbourhood-based robust-
ness, the basic idea of which is that if a small set of schedules close to the preschedule
is known to be of good quality before an unexpected event occurs, then perhaps
one of these schedules provide a good solution in case an unforeseen event occurs.
Co-evolution is a term to describe genetic algorithms in which a population of ro-
bust solutions and a population of problem instances evolve simultaneously. At the
Table 1. Single machine scheduling problem notation

Description Symbol Remarks
Number of jobs n
Release date rj
Processing time pj
Due date dj
Weight wj
Starting time tj rj ≤ tj
Completion time Cj Cj = tj + pj
Lateness status Uj Uj = 1 iff Cj > dj and 0 otherwise
end, the first population contains the most robust solutions and the second one the
most difficult problem instances. The first population might contain, e.g., a set of
schedules and the second a set of machine breakdowns.
3 Single machine scheduling problem formulation
The objective of the scheduling problem in this paper is to schedule a given set
of jobs on a single machine without preemption. Every job has a release date
that represents the time at which production or assembly of the batch can begin
at the earliest. The processing times of the jobs are known, as well as their due
dates. Depending on its importance, a weight is attached to every job. A job is
called late if its completion time is greater than its due date. The objective is to
find a schedule of jobs that minimises the total weighted number of late jobs. Late
jobs are arbitrarily moved to the end of the schedule, as the amount of lateness is
considered unimportant. Table 1 clarifies the notation
used.
The problem in this paper is denoted as 1|rj | wj Uj in the standard classifica-
tion. The problem is N P-hard in the strong sense (Lenstra et al. 1977). For the static
case, previous work on this problem can be found in Sevaux and Dauzère-Pérès
(2002), Dauzère-Pérès (1995), Dauzère-Pérès and Sevaux (2003), Dauzère-Pérès
and Sevaux (2001), Baptiste (1999) and Baptiste et al. (2003).
Although all parameters of the problem can be made stochastic, we only model
the case of stochastic release dates. We will show however, that other stochastic
parameters can easily be entered into the problem formulation and solved by the
modified GA.
4 A genetic algorithm for a single machine scheduling problem
Algorithm 1 gives a schematic representation of the GA used in the rest of the paper.
In this GA, a solution is encoded as a permutation of the n jobs. An initial
population is created by randomly generating permutations of jobs. The size of the
population N is computed by the formula N = 5+n/5 where n is the number of jobs
Algorithm 1 Basic incremental genetic algorithm

1: Generate an initial population
2: while stopping conditions are not met do
3: Select two individuals
4: Crossover the two individuals
5: Mutate offspring under probability
6: Evaluate offspring
7: Insert offspring under conditions
8: Remove an individual under conditions
9: end while
10: Report results
of the selected instance. The main loop of the GA is stopped after a fixed number
of iterations without improvement of the best solution or after a certain maximum
number of iterations. The maximum number of iterations ItMax is computed by
ItMax = 50 ∗ n ∗ N and the maximum number of iterations without improvement
is ItMaxwoImp = ItMax/10.
Two parent solutions are drawn from the population. The former is chosen
according to a ranking selection method which give a higher probability to the
best individuals (Reeves 1993) and the latter is randomly chosen. The population
is sorted in ascending order of fitness values (ascending order of the objective
function), so the best individual is ranked first. If N denotes the size of the pop-
ulation, the selection of the first parent is made using the probability distribution
2(N + 1 − k)/(N (N + 1)) where k is the k th chromosome in the ascending order
of fitness values. Using this technique, the median chromosome (N + 1)/2 has
the probability 1/N of being selected, while the best individual has the probability
2/(N + 1) (approximatively twice the median). The second parent is randomly
selected with an uniform distribution.
The crossover operator is a one-point crossover operator X1. A crossover point
is randomly chosen. The first part of parent P1 is copied into the offspring. The rest
of the chromosome is filled by reading the information of the second parent P2.
The jobs that are not already in the offspring are copied, preserving the permutation
property. See Fig. 1.
Crossover point
P1 1 2 3 4 5 6 7
P2 4 3 7 6 2 5 1 O1 1 2 3 4 7 6 5
Fig. 1. Crossover operator X1

A general pairwise interchange mutation operator permutes two jobs randomly

chosen in the chromosome. See Fig. 2. This mutation is done under a probability
Pm . For the deterministic case, we chose Pm = 0.25.
1 2 3 4 5 6 7 1 5 3 4 2 6 7
Fig. 2. Mutation operator GPI
After mutation, the solution is evaluated. The fitness value of a solution is equal
to the sum of the weights of the late jobs, i.e.,
n

f (x) = w j Uj , (1)
j =1
where Uj is defined as in Table 1.

The genetic algorithm uses incremental replacement, i.e., the population size
remains the same throughout the run. A new offspring is inserted if it improves
at least the worst solution in the population. When a new solution is inserted, a
solution randomly chosen from the worst half of the population, is discarded.
In Sevaux and Dauzère-Pérès (2002) different variations of this simple genetic
algorithm, combined with a local search technique are compared to and found to be
superior over other methods. The GA used in this paper does not attempt to locally
improve solutions using a local search operator. Experimental results show that,
when using the GA to find robust solutions, the use of a local search improvement
operator tends to overemphasise solution quality. Finding robust solutions requires
the exploration of a diverse set of solutions, covering a portion of the search space
that is as large as possible. Adding a local search operator at this step will intensify
the search and the solutions will have some jobs tightly sequenced so that any mod-
ification in the release date will cause an important modification of the optimisation
function. For our study, the local search operator will not be used.
5 A genetic algorithm for robust schedules
In the following, we describe how the genetic algorithm for the deterministic
scheduling problem can be modified so that it finds solutions that are quality robust
and/or solution robust. As in Sörenesen (2001), this is done by replacing the fitness
function of the GA by a so-called robust evaluation function. When we want to
make the distinction, the fitness function for the deterministic problem is referred
to as the ordinary evaluation function.
5.1 Quality robustness
Let x be a solution of the problem (a permutation of the jobs). The quality of x is

computed by an evaluation function f (x). When we want to indicate that f has
parameters, we write f (x, P ), where P is the set of problem data. In our case, P
represents the characteristics of the jobs (rj , pj , dj , wj ). To allow the GA to find
robust solutions, the evaluation function f (x) is replaced by a robust evaluation
function f ∗ (x). The robust evaluation function for quality robust solutions adheres
to the following principles (Sörensen 2001):
Principle 1 : Each solution is implemented on a modified set of characteristics

Si (P ). S is a sampling function, that takes a random sample from the stochastic
elements of P . Si (P ) is the i-th set of sampled parameters of P . We call the
implementation of a solution on a modified set of characteristics a derived
solution.
Principle 2 : Several evaluations of a solution x on a sample of P are combined
into a new evaluation function. An evaluation of a derived solution is called a
derived evaluation. This new function is the robust evaluation function f ∗ (x).
A possible form of a robust evaluation function is a weighted average of m

derived evaluations:
m
1
f ∗ (x) = ci f (x, Si (P )) (2)
m
i=1
where ci is a weight associated to this derived evaluation according to its importance

and m is the number of derived solutions to evaluate.
A more conservative robust evaluation function examines the worst-case per-
formance of a solution across all derived evaluations:
f ∗ (x) = max f (x, Si (P ))

i=1...m
if f has to be minimised. In this paper, we only use the robust evaluation function
of Eq. (2).
For the scheduling problem considered in this paper, we assume the release dates
to be stochastic. The robust evaluation function evaluates each solution several
times, each time on a new random instantiation of the release dates. The same
permutation of the jobs is kept but the release dates are modified. Afterwards, the
engine which computes the objective function value of the permutation is called
and executed with this new instantiation of the problem. The derived evaluations are
combined into a single robustness measure as indicated. This implies that a standard
evaluation has to be performed several times for each solution, which increases the
computational time. The advantage of this approach however is that the standard
evaluation function does not have to be modified. The implementation of a robust
evaluation function is therefore very easy and only requires a standard evaluation
function to be available and a metaheuristic already running for problems with static
data. The only change made is to the evaluation step of Algorithm 1, rendering this
way of finding robust solutions extremely easy to implement and very flexible.
5.2 Solution robustness
Solution robustness is a property of a solution that is similar to a given baseline

solution x0 , i.e. for which the distance to the baseline solution (as measured by
some distance function) is small. Of course, solution robustness cannot be used
as the only objective, since solution quality or quality robustness should always
be taken into account. The need for solution robustness therefore automatically
transforms the problem into a multi-objective one and a solution should be found
that simultaneously has a high quality (robustness) and a small distance to the
baseline solution. In our framework, solution robustness is obtained by measuring
the distance between the baseline solution and each solution generated by the GA.
It is assumed that the GA visits a sufficiently diverse set of solutions, so that at
least a fraction of them will be solution robust. A solution is then chosen using a
multi-objective decision making process – taking into account the decision maker’s
preferences for solution robustness and quality (robustness).
Edit distance. First developed in the context of correcting binary codes transmit-
ted across a faulty channel (Levenshtein 1966), the edit distance is a measure of the
similarity between two strings, composed of characters of a finite alphabet . An
extra null-character is also defined to signify the absence of a character. Three
elementary edit operations are defined (x, y ∈ ): insertion ( → x), deletion
(x → ) and substitution (x → y). An edit transformation between strings s and
t is a set of elementary edit transformations that transform s into t.
Simplifying, the edit distance d(s, t) can be defined as the number of edit
operations that is required to transform a source string s into a target string t 1
A straightforward dynamic programming algorithm to compute the edit distance
is due to Wagner and Fischer (1974). The time complexity of this algorithm is
O(|s| × |t|) where |s| and |t| are the lengths of strings s and t respectively. More
efficient algorithms have been developed (e.g. by Ukkonen 1985), but these are
beyond the scope of this paper. Under some conditions (Wagner and Fischer 1974),
the edit distance is a metric, i.e. it satisfies the conditions that (∀s, t, u): d(s, t) ≥ 0,
d(s, t) = 0 ⇔ s = t, d(s, t) = d(t, s), and d(s, t) ≤ d(s, u) + d(u, t).
1 In a more general formulation of the edit distance, a distance function γ (x, y) ≥ 0 is defined that
determines the non-negative real-valued cost of transforming character x into character y (x, y ∈ ∪).
The cost of an edit transformation is defined as the sum of the costs of the elementary edit operations it
contains. The edit distance d(s, t) now is defined as the minimum-cost edit transformation that transforms
s into t.
Distance between two schedules. If all jobs of the scheduling problem are repre-
sented by a different character, a schedule can be represented as a string of jobs.
The distance between two solutions can be calculated by the edit distance and in-
terpreted as the number of “changes” that have to be made to the first schedule to
turn it in to the second one. A “change” can be either an insertion of a job into the
schedule, a deletion of a job or the substitution of a job by another one.
A sensible distance measure should accurately reflect the “similarity” between
two schedules. The meaning of this concept is highly dependent on the specific
situation. Although the edit distance provides a relatively natural way of measuring
the similarity between two schedules (the number of changes required to transform
schedule s into schedule t), other distance measures could be developed, e.g., based
on the actual starting times of the jobs.
5.3 Incorporating risk preference
The function f ∗ (x) estimates the average performance or the worst case perfor-
mance of a solution, given that some of the parameters of the problem are stochastic.
Clearly, the worst case performance measure will lead to solutions that are more
conservative. Solutions found using this form of the robust evaluation function will
hedge only against the worst possible incidence, independent of the probability
that this scenario will occur. This type of robust evaluation function can be used by
extremely risk-averse decision makers.
A more subtle manner to incorporate the risk preference of the decision maker,
is to include into the robust evaluation function an estimate of the probability that
the quality of a solution will deviate from its expected value. A possible measure
is the standard deviation of the quality of a given solution over all samples:

m
1
∗
σ (x) = [f (x, Si (P )) − f ∗ (x)]2 .
m−1
i=1
The two measures can be integrated in a multi-objective decision making ap-

proach. A possible way is to find the solution that minimises f ∗ (x) + γ σ ∗ (x),
where γ is a parameter indicating the risk-averseness of the decision maker. A
more advanced way is to retain all efficient solutions and choose one according to
a multi-objective decision making method.
6 Application to just-in-time scheduling
6.1 Just-in-time scheduling
Many manufacturing companies have switched (at least partially) to a just-in-time

policy over the last two decades. In this philosophy, suppliers are required to deliver
their raw materials or semi-finished products to the manufacturer’s plant right on

time for these parts or materials to be taken into production (Christopher 1998). The
suppliers are required to locate a supply hub (e.g., a warehouse) close to the man-
ufacturer’s production facility and deliver parts or materials to the manufacturing
plant several times a day, at the request of the manufacturer. Based on the produc-
tion sequence, i.e., the order in which the different product batches are be produced
in the manufacturing plant, release dates and amounts to deliver of the parts or
materials are fixed by the schedulers in the manufacturing plant and communicated
to the suppliers. The production sequence determines when the production of each
batch begins, and consequently which batches will be finished on time and which
not.
Just-in-time delivery promises several advantages, the most important ones
being reduced costs and shorter lead times. A negative side-effect from the use of
just-in-time however, is the fact that any delay in the delivery by the suppliers will
postpone the production of a batch which will lead to longer and unpredictable lead
times of the finished product and increase the number of late deliveries. Although
manufacturing companies and their suppliers are working hard to reduce the delay
of parts or materials, unexpected events can have a significant effect on them. This
in turn will negatively influence the number of production batches that are finished
on time.
While manufacturers are trying hard to reduce the number of late deliveries of
parts and materials as well as the amount of time with which they are delayed, some
factors remain beyond their control. Traffic jams, truck breakdowns, strikes etc.,
are factors that may delay delivery, but are very difficult to control. To deal with
these types of unexpected events, a manufacturing company can attempt to design a
robust production sequence. A quality robust production sequence can be defined as
a production sequence, the quality of which is relatively insensitive to perturbations
in the input data. The quality of a certain schedule can be measured by various
quality measures, e.g. by counting the number of batches that are finished on time.
A solution robust production sequence may be defined as a production sequence
that is “close” to a base-line production sequence that is currently in use by the
manufacturer. As already mentioned, changing the production sequence drastically
when unexpected events occur may increase the quality but might cause a large
disruption in the regular production routine. This in turn may lead to production
errors, and other undesirable effects.
6.2 Modelling a just-in-time production system

as a scheduling problem
A just-in-time production environment can be modelled as a scheduling problem,

in which each of the production batches is represented by a job that has to be
scheduled on a single machine. Solving the problem yields a production sequence
that minimises the weighted number of late batches. Of course, other objectives
can be used, such as the weighted tardiness.
The following paragraphs describe the problem that a real French assembly
factory in the car industry faces. For reasons of confidentiality, real data were
not given, but enough information was provided to allow us to create a realistic
simulation of reality.
In the assembly factory, we observe a huge number of elementary tasks that
we aggregate in a reduced number of aggregated jobs. The number of jobs n to
be processed each day varies between 20 and 80. The jobs are quite diverse and
have widely spread-out processing times. Processing times for the different jobs are
approximately uniformly distributed, with a high variability. Jobs are planned each
day, but the schedule approximately repeats itself every week. Past observations
show that a significant part of the jobs cannot start at their expected release date
because of late deliveries of subcontractors. The percentage of jobs that is delayed
is about 20%. Jobs almost never start early. Delays are approximately uniformly
distributed between 0 and some parameter 20, where this number represents the
number of time-periods added to the original release date.
Problem instance generator – deterministic problem. From these data, we have

created a problem instance generator that generates instances according to the fol-
lowing rules. A single day of 80 five-minute periods is considered. The probability
of release dates to arise in the morning is greater than in the afternoon (and the
reverse for due dates). Hence release dates are generated according to a gamma
law ("(0.2, 4), which gives a mean of 20 and a standard deviation of 10) and due
dates are generated according to the same law but considering the horizon from
the end. If ri ≥ di the associated job is discarded. To generate only feasible jobs,
the condition di − ri ≤ 1 is checked (if not the job is discarded) and processing
times are uniformly generated in the interval [1, di − ri ]. Weights are uniformly
generated in the interval [1, 10]. In the sequel n will denote the number of jobs.
Twenty different instances are generated for each value of n and n takes its value in
the set {20, 40, 60, 80, 100}. This new set of instances will be called ODD for “One
Day Data”. Each file is numbered ODDn_i where n is the number of jobs and i the
ith instance for the specified number of jobs (ODD20_4 is the fourth instance with
20 jobs).
The rules of the problem instance generator are summarised in Table 2. U(l, u)
indicates a uniform distribution between l and u. "(a, b) indicates a gamma distri-
bution.
Robust evaluation function. In the robust evaluation function, each solution is

evaluated a fixed number of times on a modified instance of the problem data. At
each evaluation, a number of jobs is randomly chosen and the release dates of these
jobs are increased. The m evaluations are averaged to determine the value of the
robust evaluation function. The parameters of the robust evaluation function are
shown in Table 3.
Table 2. Deterministic problem instance generator rules

Parameter Value
Problem size (20,40,60,80,100)
Total time T 80
Release date rj "(0.2, 4)
Due date dj T − "(0.2, 4)
Processing time pj U (1, di − ri )
Weight wj U (1, 10)
Table 3. Robust evaluation function parameters

Parameter Value
Percentage of jobs with delayed release date 20%
Release date delay U (0, 20)
Number of evaluations per robust evaluation (m) 100
The number of evaluations m needs to be sufficiently high to avoid that a non-

robust solution is accidentally chosen. Some experiments show that a value of
m = 100 gives adequate results, while keeping the increase in computation time
within acceptable limits. The robust evaluation function value is stored and used
for population management purposes (see Sect. 4).
7 Numerical experiments
This section describes some experiments to find quality robust solutions. Experi-
ments to increase solution robustness are reported in Sect. 8.2.
All the test code has been implemented in C using a gcc version 2.96 on a Linux
RedHat 7.3 system for a Pentium III - 800Mhz PC machine.
To test the efficiency of the proposed method, a simulation procedure is used.
The standard GA, using the ordinary evaluation function, is first run on the original
data. The result obtained is called the standard sequence. The robust GA is also run
with the systematic application (i.e., for evaluation of the solution and population
management) of the robust evaluation function (Eq. 2). The result obtained is called
the robust sequence. Once these two sequences are obtained, 1000 replications of
the problem instance with randomly modified data according to the disturbances
are evaluated through the sequences. Results are analysed below. In the sequel, the
standard GA procedure will be denoted by the acronym SGA and the robust GA
by RGA.
Table 4 gives the detailed results for the SGA for the subset of instances with
80 jobs. the column ‘CPU’ gives the CPU time of the SGA in seconds. The number
of iterations is mentioned in the column ‘#It’. The next two columns respectively
show the fitness value of the best individual and the average fitness value of the
final population. The last two columns are the results of the simulation with the
1000 replications of the problem with modified instances. The sequence of the
Table 4. Results of the Standard GA for 80 jobs

Name Cpu (s) # It. Fitness Av. Pop. Av. 1000r Inc. (%)
ODD80_1 0.87 16775 400 405.05 430.17 7.54
ODD80_2 1.24 31903 349 353.90 373.00 6.88
ODD80_3 0.64 15606 348 354.19 371.25 6.68
ODD80_4 0.64 14505 411 417.14 431.00 4.87
ODD80_5 1.29 25518 307 312.76 337.28 9.86
ODD80_6 0.52 13217 329 332.62 340.99 3.64
ODD80_7 0.96 20278 331 336.86 361.88 9.33
ODD80_8 0.70 12646 354 357.67 368.84 4.19
ODD80_9 0.99 18863 317 321.43 343.06 8.22
ODD80_10 0.69 14645 344 347.67 366.58 6.56
ODD80_11 0.73 17557 394 398.19 417.80 6.04
ODD80_12 0.75 15224 363 374.52 385.68 6.25
ODD80_13 0.55 10500 317 322.81 338.57 6.81
ODD80_14 0.54 10104 364 368.86 389.80 7.09
ODD80_15 0.65 11040 369 373.86 385.32 4.42
ODD80_16 0.52 10418 370 375.90 384.52 3.92
ODD80_17 0.52 10982 325 331.48 342.11 5.26
ODD80_18 0.84 13599 307 312.81 324.60 5.73
ODD80_19 0.76 15434 357 363.43 376.29 5.40
ODD80_20 0.63 14365 365 372.33 391.88 7.36
best individual is used to evaluate these 1000 replications and the average value
is reported in the ‘Av. 1000r’ column. Column ‘Inc.’ indicates the increase (in %)
of the 1000 modified instances compared to the best solution – i.e., (Av.1000r-
Fitness)/Fitness – (e.g., for instance ODD80_1, a modification in the release dates
could lead to an increasing of 7.54 % of the objective function value, on average).
Similar results are reported in Table 5. In addition we have the robust fitness
value (column ‘Robust Fit.’) which was used during the search to guide the RGA.
This value is used as a reference instead of the Fitness value. Note that the average
population value was computed on the fitness values and not on the robust fitness
values.A column (named SGA-RGA) compares the two algorithms. The value com-
puted here is the (SGA:Av1000r - RGA:Av1000r) / SGA:Av1000r which denotes
the difference between the two algorithms in percentage after disruption.
For instance ODD80_20, for example, the best objective function value ob-
tained by the standard GA is 365 after 14365 iterations in 0.63 seconds. If we keep
this sequence knowing that release dates can be modified, we could expect an ob-
jective value of 391.88 (on average) which is an increase of 7.36% of the cost. The
Robust GA gives a sequence with a solution of 370 (or more likely with 371.78,
the robust fitness value) after 23057 iterations in 151.64 seconds. If now we use
this sequence for the modified data, the expected objective value will be 377.56 (on
average) which is an increase of only 2.04%. Although, the RGA improves almost
always the average solution after disruption of the SGA method. The solution given
by the RGA is then more quality robust.
Table 5. Results of the Robust GA for 80 jobs

Name Cpu (s) # It. Fitness Robust Fit. Av. Pop Av. 1000r Inc. (%) SGA-RGA (%)
ODD80_1 158.33 22058 406 406.36 411.76 417.60 2.86 2.92
ODD80_2 72.52 9956 357 358.91 361.90 366.73 2.72 1.68
ODD80_3 108.76 15687 369 363.37 365.62 374.23 1.42 −0.80
ODD80_4 152.49 20818 429 422.61 424.57 426.65 −0.55 1.01
ODD80_5 224.35 29705 310 314.94 317.48 328.88 6.09 2.49
ODD80_6 100.98 15109 327 328.24 335.24 335.49 2.60 1.61
ODD80_7 147.44 20249 342 345.39 344.33 358.54 4.83 0.92
ODD80_8 96.03 11853 358 360.67 363.67 365.09 1.98 1.02
ODD80_9 120.26 14271 325 327.12 330.67 337.71 3.91 1.56
ODD80_10 126.65 16910 353 355.95 359.57 358.47 1.55 2.21
ODD80_11 119.67 15989 403 408.82 414.05 419.97 4.21 −0.52
ODD80_12 94.76 13801 369 371.24 375.29 377.26 2.24 2.18
ODD80_13 62.12 8559 341 333.90 337.33 340.49 −0.15 −0.57
ODD80_14 85.73 12212 381 378.84 379.76 383.52 0.66 1.61
ODD80_15 146.25 16869 373 377.04 382.81 378.48 1.47 1.78
ODD80_16 106.09 14472 377 370.44 377.38 380.52 0.93 1.04
ODD80_17 144.83 19637 329 329.30 332.00 333.23 1.28 2.60
ODD80_18 120.86 15289 313 316.72 315.86 321.17 2.61 1.06
ODD80_19 158.79 21159 364 363.40 363.81 368.20 1.16 2.15
ODD80_20 151.64 23057 370 371.78 376.43 377.56 2.04 3.65
Table 6. Global SGA and RGA results (deviation and CPU time)
Number Gap to best solution Gap (in %) CPU time (s)
of jobs Std GA (%) Rob GA (%) SGA-RGA Std GA Rob GA
20 11.95 3.84 2.88 0.03 1.73
40 8.47 2.89 1.61 0.13 10.27
60 7.35 3.08 0.89 0.36 45.81
80 6.30 2.19 1.48 0.75 124.93
100 5.32 2.01 0.47 1.53 215.69
Table 6 gives the summarised results for the whole set of instances from 20 to
100 jobs. The columns “Gap to the best solution” indicates the average value in
percentage between the sequence found by any method and the results after the 1000
replications of modified instances (e.g., for n = 80, the average value of column
“Inc.” in Table 4 and 5). The middle column “Gap SGA-RGA” is the average value
of the column SGA-RGA of Table 5 and measures the improvement of the RGA
over the SGA after disruption.
The results confirm that it is always better to take the stochastic nature of some of
the data into account when determining a schedule. On average, modifications of the
problem data have a much smaller effect on the objective function value of solutions
found by the RGA. However, since the computation of the robust evaluation function
requires the ordinary evaluation function to be used 100 times, CPU times are much
larger for the RGA. A possible strategy to lower the computational effort is to use
the robust evaluation function only when the standard evaluation of a solution is
100
+
95 +
+
Robust evaluation f ∗ (x)
90 +
+
+
rs
rs
85
++
80 + +
rs
rs
+ +
++ + + +
75 + ++++++ + + + +
+ ++ ++ + +
+++++++++++++++++++++++++++ + +++
rs sr rs
++++
+++
+++++++++++ +
+ +
+
+
++
+
+++++
++ +
++++
++
++
+++
++
+
+
++
+
+
++
+ ++
++
+ +++++
++
++
+
+
+
++
+++
+++
+
+
++
+
+
++
+ ++ ++
++++++
+ + + + + +++
++++
++++
++++
+
+ + + + ++
rs
+++ ++++++ ++
+++ ++
++ +++++
+
++ ++
++++++++++
++ ++++
+++
+++++++++++
+++++++ ++++++++ +++ + + +
rs
70
0 10 20 30 40 50 60
Standard deviation σ ∗ (x)
Initial solutions All solutions +

rs
Fig. 3. Standard deviation of the robust value – all solutions
better than a certain threshold. We did not implement such strategies here since the
computational times recorded for the RGA are still very reasonable.
8 Helping the decision maker
8.1 Risk analysis
As mentioned in Sect. 5.3, choosing a robust solution is not an easy task and can
be done through a multiple objective decision making process. In Fig. 3, for a 20
job instance, we plot all the solutions (including the initial solutions) in the space
“Standard Deviation of the robust value / Robust Value”. The initial solutions are
spread out in the space and all the other solutions seem to converge toward a good
robust fitness value (to the bottom of the graph). At the same time, the standard
deviation of the solutions decreases meaning that the solution are more and more
robust. Figure 4 shows only the non-dominated solutions (Pareto solutions) that
can be suggested to the decision maker.
8.2 Solution robustness
To obtain solutions that are solution robust, the edit distance between a base-line
solution and all the solutions found by the RGA is measured. We assume that the
71.05
rs
71
70.95
rs
rs
70.9
rs
70.85 rs
70.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1
Standard deviation σ ∗ (x)
Pareto solutions
rs
Fig. 4. Standard deviation of the robust value – efficient set
solution that is currently implemented (the base-line solution x0 ) is the best solution
obtained with the SGA. This choice is arbitrary and in real-life situations, x0 would
represent the solution actually used. For the same 20-job instance, Fig. 5 represents
the initial, intermediate and final solutions obtained by the RGA. The number of
intermediate solutions can be very large and only non-dominated solutions are kept
for clarity.
For this particular instance, the minimum distance to x0 that can be obtained by
the GA is 12, indicating that all robust solutions are relatively “far” from the base-
line solution. Using the information about the distance to the base-line solution and
the robustness of a solution, the decision maker can make an informed decision.
9 Conclusion
In this paper, a genetic algorithm for a single machine scheduling problem was
modified to find robust solutions. Two types of robustness were defined and it was
shown how our GA can be used to find solutions that are both solution robust and
quality robust. It was also shown how information about the risk associated with
a specific solution can be incorporated into the decision making process. Our GA
was applied to a just-in-time production planning problem. Computational results
show that robust solutions can be found by applying some simple modifications
of a genetic algorithm for a deterministic problem. This approach is very flexible
120
rs
115
110
105
rs
100
95
rs
rs
rs
rs
rs sr
rsrs
90 rs
rs rs
+ +
rs
+ srrs
85 + +rs
+ +rs + +
80
12 13 14 15 16 17 18 19 20
Edit distance to baseline d(x, x0 )
Initial population Final population rs

rs
Intermediate solutions +
Fig. 5. Distance between solutions
in that it does not impose any requirements on the number and type of stochastic
information that is incorporated. Without changing the solution algorithm in an
extensive way, any information that is available about the problem data can be used
in the decision making process.
Acknowledgement. The authors would like to thank the anonymous referees and the editor for their
constructive comments and suggestions.
References
Baptiste Ph (1999) An O(n4 ) algorithm for preemptive scheduling of a single machine to minimize the
number of late jobs. Operations Research Letters 24: 175–180
Baptiste Ph, Péridy L, Pinson E (2003) A branch and bound to minimize the number of late jobs on
a single machine with release time constraints. European Journal of Operational Research 144:
1–11
Branke J (1998) Creating robust solutions by means of evolutionary algorithms. In: Parallel Problem
Solving from Nature V (LNCS, vol. 1498). Springer, Berlin Heidelberg New York, pp 119–128
Branke J (2001) Reducing the sampling variance when searching for robust solutions. In: Spector L
et al. (ed), GECCO 2001 – Proceedings of the genetic and evolutionary computation conference.
Morgan Kaufmann Publishers, pp 235–242
Christopher M (1998) Logistics and supply chain management – strategies for reducing cost and im-
proving service. Financial Times Prentice Hall, London
Dauzère-Pérès S (1995) Minimizing late jobs in the general one machine scheduling problem. European
Journal of Operational Research 81: 134–142
Dauzère-Pérès S, Sevaux M (2003) Using lagrangean relaxation to minimize the weighted number of
late jobs. Naval Research Logistics 50 (3): 273–288
Dauzère-Pérès S, Sevaux M (2001) An exact method to minimize the number of late jobs in single
machine scheduling. Technical report, Ecole des Mines de Nantes, France, submitted
Davenport AJ, Beck JC (2000) A survey of techniques for scheduling with uncertainty. Unpubl.
manuscript
Goldberg DE (1989) Genetic algorithms in search, optimization and machine learning. Addison Wesley,
♣ ((insert publ.’s place))
Hart E, Ross P (1999a) The evolution and analysis of a potential antibody library for job-shop scheduling.
In: Corne D, Dorigo M, Glover F (eds) New ideas in optimization. London, McGraw Hill, pp 185–
202
Hart E, Ross P (1999b) An immune system approach to scheduling in changing environments. In:
GECCO 99. Proceedings of the genetic and evolutionary computation conference. Morgan Kauf-
mann publishers, ♣, pp 1559–1566
Herroelen W, Leus R (2002) Project scheduling under uncertainty – survey and research potentials.
Invited paper to be published in the special issue of EJOR containing selected papers from PMS2002
Holland JH (1975) Adaptation in natural and artificial systems. Technical report, University of Michigan,
Ann Arbor
Jensen MT Robust and flexible scheduling with evolutionary computation. PhD thesis, University of
Aarhus, Dept. of Computer Science, Denmark
Kouvelis P, Yu G (1997) Robust discrete optimisation and its applications. (Nonconvex Optimization
and its Applications, vol. 14). Kluwer Academic Publishers, Dordrecht
Lenstra JK, Rinnooy Kan AHG , Brucker P (1977) Complexity of machine scheduling problems. Annals
of Discrete Mathematics 1: 343–362
LeonVJ, Wu SD, Storer RH (1993) Robust scheduling and game-theoretic control for short-term schedul-
ing of job-shops. In: Fandel G, Gulledge T, Jones A (eds) Operations research in production plan-
ning and control. Springer, Berlin Heidelberg New York, pp 321–335
Leon VJ, Wu SD, Storer RH (1994a) A game-theoretic control approach for job shops in the presence
of disruptions. International Journal of Production Research 32 (6): 1451–1476
Leon VJ, Wu SD, Storer RH (1994b) Robustness measures and robust scheduling for job shops. IIE
Transactions 26: 32–43
Levenshtein VI (1966) Binary codes capable of correcting deletions, insertions, and reversals. Soviet
Physics – Doklady 10: 707–710
Portmann MC (1996) Genetic algorithm and scheduling: a state of the art and some propositions. In:
Proceedings of the workshop on production planning and control. Mons, Belgium, pp I–XIV
Reeves CR (ed) (1993) Modern heuristic techniques for combinatorial problems. Wiley, New York
Reeves CR (1997) Genetic algorithms for the operations researcher. Informs Journal on Computing 9
(3): 231–250
Sevaux M, Dauzère-Pérès S (2002) Genetic algorithms to minimize the weighted number of late jobs
on a single machine. European Journal of Operational Research (forthcoming)
Sörensen K (2001) Tabu searching for robust solutions. In: Proceedings of the 4th metaheuristics
international conference. Porto, Portugal, pp 707–712
Tsutsui S (1999) A comparative study on the effects of adding perturbations to phenotypic parameters
in genetic algorithms with a robust solution searching scheme. In: Proceedings of the 1999 IEEE
systems, man, and cybernetics conference (SMC’99 Tokyo). pp III–585–591
Tsutsui S, Ghosh A (1997) Genetic algorithms with a robust solution searching scheme. IEEE Transac-
tions on Evolutionary Computation 1: 201–208
Tsutsui S, Ghosh A, Fujimoto Y (1996) A robust solution searching scheme in genetic search. In: Voigt
HM, Ebeling W, Rechenberg I, Schwefel HP (eds) Parallel problem solving from nature – PPSN
IV, vol. 10. Springer, Berlin Heidelberg New York, pp 543–552,
Tsutsui S, Jain JC (1998) Properties of robust solution searching in multi-dimensional space with genetic
algorithms. In: Proceedings of the 2nd International Conference on knowledge-based electronic
systems (KES-98).
Ukkonen E (1985) Finding approximate patterns in strings. Journal of Algorithms 6: 132–137
Wagner RA, Fischer MJ (1974) The string-to-string correction problem. Journal of the ACM 21: 168–173
GA|PM: genetic algorithms with population
management
Kenneth Sörensen and Marc Sevaux
February 2004
Abstract
We present a new type of evolutionary algorithms called genetic
algorithms with population management or GA|PM. The main distin-
guishing feature of GA|PM is the use of distance measures to control
the diversity of a small population. Combined with a local search pro-
cedure to ensure the high quality of individual solutions in the pop-
ulation, population management strategies can be developed to avoid
slow or premature convergence and achieve excellent performance.
The new algorithm is tested on two problems: the multidimensional
knapsack problem and the weighted tardiness single machine schedul-
ing problem. On both problems, population management is shown to
be able to improve the performance of an evolutionary algorithm.
1 Introduction
“Classical” genetic algorithms (GA) —pioneered by among others Holland
[15] and Goldberg [12]— have been shown experimentally to perform rather
poorly on many combinatorial optimization problems. The performance of
many GA implementations is characterized by large memory requirements,
large running times and poor performance in terms of solution quality.
Commonly referred to as evolutionary algorithms (EA), many variants of
genetic algorithms have been proposed to improve their performance. Per-
haps most successful in combinatorial optimization are memetic algorithms
[20], also called hybrid genetic algorithms or genetic local search. These al-
gorithms apply local search operators to improve the quality of individual
solutions in the population. Besides hybridizing genetic algorithms, the use
of more natural encodings (instead of binary ones) is also widely recognized
as contributing to the quality of an EA, as is the use of crossover opera-
tors that preserve the structure of good parts of solutions (called building
blocks). That this is not a trivial task is exemplified by van Kampen and
Buydens [30], who show that their genetic algorithm is easily outperformed
by a simulated annealing approach because the crossover operator is unable
to preserve the fitness of high-quality building blocks.
1
Nevertheless, even EA with all these improvements may still fall victim
to either slow or premature convergence. Many researchers agree that the
quality of a metaheuristic optimization approach is largely a result of the
interplay between intensification and diversification strategies (see e.g. Fer-
land et al. [7], Laguna et al. [16]). One of the main motivations for this
paper is the observation that the design of evolutionary algorithms makes
it particularly difficult to control the balance between intensification and
diversification. As Hertz and Widmer [14] point out, preserving the diver-
sity of the population of an evolutionary algorithm is necessary. Although
EA have the operators to increase or decrease the diversity of the popula-
tion, most lack the means to control this diversification. On the one hand,
crossover and mutation operators tend to diversify the population by intro-
ducing new or recombining existing genetic material. As a result of too much
diversification, some EA exhibit extremely slow convergence. On the other
hand, the hybridization (local search) operators tend to focus the search in
specific regions. This may even lead to so-called premature convergence: a
total loss of diversity in the population.
This problem is tackled by a new class of evolutionary algorithms that we
propose: GA|PM or genetic algorithms with population management [28].
Population management works by measuring and controlling the diversity of
a small population of high-quality solutions, and makes ample use of distance
measures. The distance measure to use depends on the representation of a
solution. For binary problems, the Hamming distance can be used. For
permutation problems, several distance measures have been proposed in the
literature, e.g. [2, 25, 26]. The distance measures used in this paper are
based on the edit distance (see e.g. Wagner and Fischer [31]), that is well
known and easily adaptable to a large number of combinatorial problems
[27].
Working with a small population (Greistorfer and Voß [13] call it a pool)
is an idea that is incorporated in a number of metaheuristic approaches, but
a common framework for creating and maintaining such a population has not
yet been established. Scatter search and path relinking [9] are approaches
that work by combining solutions from a small set of high-quality (called
elite) solutions. As mentioned, memetic algorithms [20] combine genetic
algorithms with local search to obtain a small population of high- quality
solutions. Some work on standard genetic algorithms using a small popula-
tion is due to Reeves [23].
GA|PM show some resemblance to memetic algorithms [20] but they
differ from them in that GA|PM use population management. Other ap-
proaches that use a small population of high quality solutions include scatter
search and path relinking [9]. GA|PM offer the advantage of being closer
to classical EA in terms of algorithm structure and therefore considerably
easier to implement. The use of distance measure for GA design has been
proposed for maintaining a diverse population [19] or to locate a set of dif-
2
ferent solutions of a multimodal problem such as in crowding [18] or fitness
sharing [11]. GA|PM differ from these approaches in that they maintain a
small population of locally improved solutions and use adaptive population
management strategies.
The rest of this paper is structured as follows. Section 2 describes the ba-
sic structure of a GA|PM and discusses how distance measures can be used
to control the diversity of a population. Some possible population manage-
ment strategies are also discussed. In section 3, the algorithm is tested on the
multidimensional knapsack problem and the total weighted tardiness single-
machine scheduling problems. For both problems, we compare the GA|PM
to a hybrid evolutionary algorithm without population management. Al-
though the population management improves the performance of the EA
in a big way in both cases, we should remark that the main goal of these
applications was to show the effectiveness of population management, not
to create solution methods that could compete with the best-performing
approaches. That GA|PM can be used to create first-class optimization
methods, is currently being shown in ongoing research. In Sörensen [28], a
GA|PM for the vehicle routing problem (VRP) is shown to produce results
competitive to the best-known approaches. In Prins et al. [22], a GA|PM
is developed for the capacitated arc routing problem (CARP). To date, this
GA|PM is the best-known approach for this problem.
2 GA|PM
This section describes the new genetic algorithm with population manage-
ment. Its main distinguishing features are
• a small population (typically 10 to 30 solutions),
• a local improvement operator (e.g. a local search, or simple tabu search

procedure),
• population management to control the diversity of the population.
2.1 Algorithm overview

A genetic algorithm with population management is structured much like a
standard genetic algorithm, but differs in the use of population management
and local search. An outline is given in algorithm 1.
Initially, a small population is built randomly or using initial heuristics.
From this population, two parent solutions are selected and subjected to the
crossover operator, forming one or two new offspring solutions. These solu-
tions are improved by a local search operator and added to the population,
after being subjected to population management.
3
Algorithm 1 GA|PM outline
1: initialize population P
2: set population diversity parameter ∆
3: repeat
4: select: p1 and p2 from P
5: crossover: p1 ⊗ p2 → o1 , o2
6: local search: improve o1 and o2
7: for each offspring o do
8: while o does not satisfy conditions for addition (input function)
do
9: mutate o
10: end while
11: remove solution: P ← P \b
12: add solution: P ← P ∪ o
13: end for
14: update diversity parameter ∆
2.2 Population management

Population management controls the diversity of a small population of high-
quality solutions. It uses an input function that determines whether an
offspring solution is added to the population or not. In GA|PM, this input
function takes the following two factors into account:
• the quality of the solution, and
• the diversity of the population after addition of the solution, or —in

other words— the contribution that the solution makes to the diversity
of the population; this is measured as the “distance” of the solution
to the population, see section 2.2.2.
2.2.1 Distance measures

To evaluate whether a candidate solution sufficiently diversifies the popula-
tion, a distance measure d is used that determines for each pair of solutions
their relative distance (or similarity). The distance should be measured
in the solution space and not —as is commonly done— in the objective
function space. Distance measures cannot be developed independent of the
problem or even the representation (encoding) of the solution. For binary
representations, the Hamming distance can be used. For solutions that are
encoded as vectors of real numbers, some form of the Minkowsky-r-distance 1
n
!1
X r
1 r
y) =
d(~x, ~ |xi − xj |
i=1
4
is appropriate (e.g. Euclidean, Manhattan, Chebychev).
For permutation problems, several distance measures have been devel-
oped in the literature. A GA|PM developed in this paper (see section 3.2.1)
uses the so-called edit distance. The edit distance can be calculated between
two strings composed of characters from a finite alphabet. The edit distance
is the number of edit operations required to transform the first string into
the second one2 . Three edit operations are defined: insertion of a character,
deletion of a character and substitution of a character by another one. Us-
ing a simple dynamic programming algorithm [31], the edit distance can be
calculated in O(n2 ). Other, more efficient algorithms have been developed,
e.g. [29]. The edit distance can be modified for different types of permu-
tation problems, where solutions cannot be represented as a simple string.
An example is the traveling salesman problem, in which a solution does not
have a fixed starting position. For a more elaborate discussion, we refer to
Sörensen [27].
2.2.2 Distance of a solution to the population

Given a distance measure that can calculate the distance between any pair
of solutions, the distance of a given solution sk to the population can be
calculated as follows:
dP (sk ) = min d(sk , si ). (1)
si ∈P
Calculating the distance of a solution to the population requires calcu-

lating |P | distance measures (where |P | is the cardinality of the population).
This high computational requirement is one of the reasons why population
management techniques are more effective when applied to small popula-
tions.
2.2.3 Input function and diversity parameter

It is obvious that a solution that has a small distance to another solution
already in the population, will not contribute much to the diversity of a pop-
ulation. Therefore, a solution is not added to the population if its distance
to the population is below a certain threshold ∆. We call ∆ the diversity
parameter. Assuming that the quality of sk is sufficient, a solution can be
added to the population if the following holds:
dP (sk ) = min d(sk , si ) ≥ ∆. (2)

si ∈P
Using the distance dP (sk ) and the fitness or objective function value
f (sk ), the input function can also use a multi-objective decision method to
2
Often, a cost is assigned to each possible edit operation and the edit distance is defined
as the minimum total cost of all edit operations required to transform the first string into
the second.
5
determine whether sk should be added to the population or not. A very
simple way of doing this (assuming that f should be minimized) is to calcu-
late f (sk ) + λdP (sk ). If this value does not exceed a certain threshold, the
solution is added, otherwise it is discarded. λ is a parameter that determines
the relative importance of the diversity with respect to the solution quality.
Of course, more elaborate multi-objective decision methods can be used.
If the local search procedure is effective enough to always ensure the
quality of solutions it produces, a solution can be added if eq. (2) holds,
without taking the objective function value of the solution into account.
As shown in algorithm 1, a solution that does not have a sufficiently large
distance to the population, is randomly mutated until it does. Of course,
other strategies are possible, such as simply discarding the solution.
2.2.4 Population management strategies

Using the diversity parameter ∆, the diversity of the population can be
controlled as higher values for ∆ will increase the diversity of the population
while lower values will decrease it. A high value of ∆ will allow only solutions
that have a large distance to all solutions in the population and will lead
—perhaps after a few iterations— to a population that consists of very
different solutions. A low value of ∆ will allow solutions in the population
that are relatively similar to solutions already in the population. This will
result in a less diverse population.
Several population management strategies can be proposed, using only
the diversity parameter ∆. The following lists several potential strategies in
increasing order of complexity.
Strategy 1: ∆ is set to a constant level. This strategy prevents population

convergence, and introduces a constant level of diversification into the
population.
Strategy 2: ∆ is set to a high level in the beginning of the algorithm.

The value of ∆ is decreased steadily throughout the working of the
algorithm, allowing increased intensification near the end of the run.
Strategy 3: A closely related alternative strategy is to create a few good

solutions (low ∆) in the beginning of the algorithm and allow more
diversification near the end. A variant of this strategy is used in the
experiments.
Strategy 4: ∆ is set to a high level in the beginning of the algorithm. It

is steadily decreased as better solutions are found. When no improve-
ment of the best solution is found for a set number of iterations, the
diversity is increased by increasing ∆, thus introducing new genetic
material in the population. After this, ∆ is steadily decreased again,
6
∆ ∆
Generations Generations
(a) Strategy 1 (b) Strategy 2
No improvements found
∆ ∆
Generations Generations
(c) Strategy 3 (d) Strategy 4
Figure 1: Population management strategies
etcetera. This strategy can be called adaptive because it uses informa-

tion about the effectiveness of the search to dynamically control the
diversity of the population.
These strategies are graphically represented in figure 1. A higher com-

plexity of a population management strategy also implies an increase in the
number of parameters that has to be determined. Elaborate strategies like
3 and 4 are especially useful when the search space is large and algorithm
runs are long.
2.3 Selection, crossover and other genetic operators

Several genetic operators have to be defined in order for the algorithm to
work. The selection operator —used for the selection of parent solutions and
for the selection of solutions to remove from the population when new solu-
tions need to be added— can be a binary tournament selection, a roulette
wheel selection or any other selection method.
Crossover operators should be designed to preserve the good features of
the parent solutions as much as possible. A bad choice of crossover operator
will result in offspring solutions that most likely have a rather poor quality.
7
For the design of efficient selection and crossover operators, we refer to
the specialized literature. See e.g. Reeves [24] for a discussion.
2.4 Intensification
A local search procedure is necessary to maintain a population of high-
quality solutions. This local search procedure should be able to quickly
improve the quality of a solution produced by the crossover operator, with-
out diversifying it into other regions of the search space. Neighbourhood
-search methods like simple tabu search approaches are particularly useful
for this purpose.
3 Experiments
In this section, we apply the principles of GA|PM to two problems: the
multidimensional knapsack problem and the total weighted tardiness sin-
gle machine scheduling problem. We should note that in none of the two
problems, it is our intention to compete with the best approaches in the
literature as this would involve developing specialized crossover operators
and advanced local search operators and would distract the attention from
the main point: the population management of GA|PM. Instead, we fo-
cus on the population management and compare the GA|PM we develop to
comparable EA that lack population management.
3.1 The multidimensional knapsack problem

The N P-hard 0/1 multidimensional knapsack problem (MKP01) is a pure
0/1 integer programming problem. Given a set of n items each having an
associated profit ci , the objective of this problem is to select the subset of
items that maximizes the total profit, while satisfying a set of knapsack
constraints. For each knapsack j, each item i has a given weight aij and the
sum of the weights of the chosen items is not allowed to exceed the knapsack
capacity bj . More formally, the MKP01 can be stated as follows:
X
n
max ci x i
i=1
Xm
(3)
s.t. aij xi ≤ bj ∀j ∈ [1, m]
i=1
xi ∈ 0, 1 ∀i ∈ [1, n]
Several approaches have been proposed for this problem, including ge-
netic algorithms [3] and tabu search [6, 10].
8
3.1.1 A GA|PM for the MKP01
A solution of the MKP01 can be represented as a binary string of length n.
We set the Pobjective function value of a feasible solution equal to the total
profit (i.e. i ci xi ). For infeasible solutions, we set the objective function
value equal to the (negative) total violation of all constraints combined, i.e.
X X
(bj − aij xi ) (4)
j i
Since feasible solutions always have a positive evaluation function value and
infeasible solutions a negative one, a feasible solution is always preferred and
the search will not return an infeasible solution unless no feasible solution
could be found. Assigning a negative objective function value to an infeasible
solution allows us to use the local search procedure to improve such solutions,
assuming that the quality of an infeasible solution is determined by the total
violation of the constraints.
A simple steepest descent local search procedure is used. This procedure
attempts to change the status of every item of a given solution and changes
the status of the item that yields the largest increase in objective function
value. The search continues until the objective function value cannot be
increased by changing the status of a single item.
The distance between two solutions is measured as the Hamming distance
between their binary strings. As usual, the distance of a solution to the
population is given by the minimum distance of this solution to any solution
already in the population.
Selection is done by a binary tournament method: two solutions are
randomly chosen and the best one is selected. A simple one-point crossover
is used to combine solutions. Each crossover produces two offspring solu-
tions that are locally improved using the local search procedure and then
subjected to population management.
Population management determines the distance dP (s) of a given solu-
tion to the population and compares this value to the population diversity
parameter ∆. If this distance is smaller than ∆, the solution s is mutated.
This is done by randomly flipping ∆ − dP (s) bits of s. This procedure does
not ensure that the mutated solution has a sufficiently large distance to the
population, but it diversifies it without causing too much overhead.
The performance of the GA|PM is compared to the performance of the
evolutionary algorithm that results from removing the population manage-
ment component from the GA|PM and replacing it with random mutation.
From now on, we refer to this procedure as HGA (hybrid GA).
9
3.1.2 Experiments
The GA|PM is tested on 48 well-known test problems available from the OR
library3 .
A first experiment uses the simplest population management for the
GA|PM, i.e. ∆ = 1. A second experiment uses more advanced population
management (a variant of strategy 3 in figure 1) with
current iteration n
∆= × ,
maximum iterations 5
where n is the number of items. This strategy sets ∆ to 0 in the beginning
of the algorithm and linearly increases it to a maximum of n/5.
A population of 10 is used in all experiments. The HGA uses a 10%
random mutation rate. Each algorithm is awarded 10000 generations to
find the optimal solution. Time is measured when the optimal solution is
found or at the end of the 10000 generations when it is not found. Each
experiment is repeated 5 times.
3.1.3 Numerical comparison

Results show that the GA|PM outperforms the HGA. The average quality
over the 5 runs of the solutions produced by the GA|PM is better in all
cases. In 46 out of 55 cases (84%), the worst solution found by the GA|PM
is better than the best solution found by the HGA. More detailed results
can be found in table 1.
Nr. opt. Min. 1 opt. Avg. it. Avg. CPU
GA|PM strategy 1 3.84 87.27% 3537 6.43
GA|PM strategy 3 3.73 92.59% 3879 10.00
HGA 0.51 25.45% 9206 9.38
Table 1: Comparison GA|PM and HGA for the multidimensional knapsack

problem
This table should be read as follows. “Nr. opt.” is the average number
of times the optimal solution was found over the 5 runs, “Min. 1 opt.” is the
average percentage of instances for which the optimal solution was found at
least in one of the 5 runs, “Avg. it.” is the average number of iterations,
and “Avg. CPU (s)” is the average time required in seconds.
As can be seen, the GA|PM performs much better than the HGA without
population management. The average time per iteration is somewhat higher,
but the solutions found are much better and the GA|PM therefore performs
less iterations. The more advanced population management increases the
average number of iterations required and also the average CPU time. It
3
http://www.ms.ic.ac.uk/jeb/pub/mknap2.txt
10
does however improve the robustness of the algorithm in that the optimal
solution is now found in over 92% of the problem instances.
3.2 The total weighted tardiness single machine scheduling

problem
A set of n jobs has to be sequenced on a single machine. Preemption is not
allowed. Jobs are not available before a release date rj and are processed
for pj units of time. For each job, a due date dj and a weight wj is given.
If Cj denotes the completion time of job j, the tardiness is defined by Tj =
max(0, PCj − dj ). The objective is to minimize the total weighted tardiness
(1|rj | wj Tj ). This problem is N P-Hard in a strong sense.
The total weighted tardiness problem with release dates is probably one
of the most difficult
Pone machine scheduling problem. If all the weights are
equal to one (1|rj | Tj ) the problem is N P-Hard in a strong Psense [8]. For
different weights and the release dates all equal to zero (1| | wj Tj ), the
problem is also N P-Hard in a strongP sense [17].
Recent approaches on the 1| | wj Tj problem are proposed in the lit-
erature. In [5], a comparison between simulated annealing, tabu search and
genetic algorithms is discussed and in [4] the iterated dynasearch algorithm
is presented and shown to yield the best solutions. Results are reported in
the OR-Library. A recent constraint-based approach [1] can solve problem
instances with up to 40 jobs.
3.2.1 A GA|PM for the total weighted tardiness problem

In this section, we develop a simple GA|PM for the total weighted tardi-
ness problem and compare it to a similar hybrid GA that lacks population
management. To perform a fair comparison between the two approaches,
common components are used for the HGA and GA|PM algorithms.
The selection process used is a binary tournament selection, the crossover
operator is the standard LOX operator [21] and the mutation operator is the
GPI (general pairwise interchange). The local search procedure is also based
on the GPI and a steepest descent algorithm is applied. The algorithm is
stopped after 60 seconds of CPU time.
For the numerical experiments, we use a variation of strategy 3. The
distance measure used is the edit distance discussed in section 2.2.1. In the
implementation, ∆ is measured in percentage (result of the distance measure
divided by the maximum distance value). At the initialization step, ∆ is
fixed to 1%. Every 15 iterations, ∆ is multiplied by a growth factor (1.05).
When an upper limit of 50% is reached, ∆ is not increased anymore, since
this would not allow the algorithm to find a new solution that is accepted
during the search.
11
In the HGA, not all solutions are subjected to local search as this almost
always induces premature convergence. Instead, the local search rate p ls is
set to 10%, the mutation rate pm is fixed to 10%. The size of the population
is set to 25. For the GA|PM, the size of the population is only 10 individuals.
3.2.2 Numerical comparison

Results are reported in different tables. Table 2 summarizes the results
for the ODD instances (these instances can be obtained by email request).
Optimal solutions are not known for this set. The first column indicates the
set of instances (ODD20 is the ODD instances with 20 jobs), the following
three columns provides the results for the GA|PM method and the next three
columns, the results for the HGA method. Note that for a fair comparison,
the same amount of time (60s) has been given to the GA|PM and the HGA.
GAPM results HGA results

Set of First Avg. gap Avg. First Avg. gap Avg.
inst. pos. (in %) iter. pos. (in %) iter.
ODD20 0 0.000 32375 0 0.000 952757
ODD40 0 0.000 4142 0 0.000 258526
ODD60 6 0.088 1685 1 0.251 88106
ODD80 6 0.003 1153 1 0.083 39040
ODD100 8 0.064 844 4 0.118 19854
Global 20 0.057 8040 6 0.148 271657
Table 2: Numerical results for ODD instances, CPU time = 60s
For each of the methods, the first column gives the number of times
the method reports the best solution (among the two methods), the second
column is the average deviation from the best solution. The last column
gives the average number of iterations done in 60 seconds.
P
The OR-Library instances are generated for the 1| | wj Tj problem and
can be solved by our approach. Optimal values of solutions are available for
124 and 115 out of 125 of the 40 and 50 job problem instances, respectively.
When the optimal solution is unknown, best solutions provided by [5] are
used instead (by extension they are also called optimal solutions).
12
ORLib40 results
Method Opt. First Avg. Gap Iterations
used sol. pos. (in %) Avg. Max
GAPM 125 4 0.000 5682 8961
HGA 121 0 0.284 327763 470984
ORLib50 results
Method Opt. First Av. Gap Iterations
GAPM 123 11 0.002 3334 5298
HGA 113 0 0.595 196024 287802
ORLib100 results
Method Best. First Av. Gap Iterations
GAPM 94 27 0.276 1030 1480
HGA 85 16 2.110 34164 56521
Table 3: Numerical results for OR-Library instances, CPU time = 60s
In table 3 the column “Opt. sol.” (“Best sol.” for 100 job instances)
counts the number of times the algorithm find the optimal solution (or best
known). The rest of the columns is identical to the previous table except
that the deviation is measured from the optimal solution. The maximum
number of iterations is added too. Again, the same amount of time has been
allocated to the two methods for a fair comparison.
These results show that the GA|PM approach performs better than the
HGA approach, even though it operates on a smaller population. For the
OR-Library instances, more optimal solutions are found and the approach
is ranked first in more cases. Moreover, the deviation from the optimal (or
best known) solution is always smaller with the proposed method. When
we use the same stopping conditions (fixed number of iterations without
improvement of the best solution) the difference between the two algorithms
is even greater, but the GA|PM requires more time.
4 Conclusions
It is commonly known that the performance of an evolutionary algorithm
(and any metaheuristic) on a hard combinatorial optimization problem de-
pends in a big way on the balance between intensification and diversification.
Although all evolutionary algorithms possess the necessary operators for in-
tensification and diversification, many EA implementations lack a mecha-
nism to control the balance between these two factors. This paper intro-
duced a new class of evolutionary algorithms, GA|PM or genetic algorithms
with population management, that addresses this issue. GA|PM are hy-
brid evolutionary algorithms that —like scatter search— operate on a small
population of high-quality individuals. The main distinguishing feature of
13
a GA|PM is the application of population management strategies to control
the diversity of the population. We discussed how such strategies can be
easily implemented by controlling the value of only a single parameter, the
population diversity parameter ∆.
The proposed approach was tested by applying it to two different com-
binatorial problems: the multidimensional knapsack problem (a pure 0/1
integer programming problem) and the total weighted tardiness single ma-
chine scheduling problem, a difficult permutation problem. We discussed
that different problems require different distance measures and proposed
to use the edit distance for permutation problems. For binary problems,
the Hamming distance was used. Numerical comparison showed that the
GA|PM outperformed very similar hybrid genetic algorithms without pop-
ulation management (but with random mutation instead).
References
[1] P. Baptiste, A. Jouglet, and J. Carlier. A branch-and-bound procedure
to minimize total tardiness on one machine with arbitrary release dates.
European Journal of Operational Research, 2004. To appear.
[2] V. Campos, M. Laguna, and R. Martı́. Context-independent scatter and

tabu search for permutation problems. INFORMS Journal on Comput-
ing, to appear, 2003. URL http://leeds.colorado.edu/Faculty/
Laguna/publications.htm.
[3] P.C. Chu and J.E. Beasley. A genetic algorithm for the multidimen-
sional knapsack problem. Journal of Heuristic, 4:63–86, 1998.
[4] R. Congram, C.N. Potts, and S. van de Velde. An iterated dynasearch

algorithm for the single-machine total weighted tardiness schedulign
problem. Informs journal on computing, 14(1):52–67, 2002.
[5] H.A.J. Crauwels, C.N. Potts, and L.N. Van Wassenhove. Local search
heuristics for the single machine total weighted tardiness scheduling
problem. Informs Journal of Computing, 10(3):341–350, 1998.
[6] F. Dammeyer and S. Voß. Dynamic tabu list management using the
reverse elimination method. Annals of Operations Research, 41:31–46,
1993.
[7] J.A. Ferland, S. Ichoua, A. Lavoie, and E. Gagné. Scheduling using tabu
search methods with intensification and diversification. Computers and
Operations Research, 28:1075–1092, 2001.
[8] M.R. Garey and D.S. Johnson. Computers and intractability: a guide
to theory of np-completeness. Freeman, San Franscisco, USA, 1979.
14
[9] F. Glover. A template for scatter search and path relinking. In J.-
K. Hao, E. Lutton, E. Ronald, M. Schoenauer, and D. Snyers, editors,
Artificial Evolution, volume 1363 of Lecture Notes in Computer Science,
pages 13–54, Springer, Berlin, 1998.
[10] F. Glover and G.A. Kochenberger. Critical event tabu search for the
multidimensional knapsack problem. In I.H. Osman and J.P. Kelly,
editors, Metaheuristics: The Theory and Applications, pages 407–427.
Kluwer Academic Publishers, Boston, 1996.
[11] D. Goldberg. Genetic algorithms with sharing for multimodal function

optimization. In Proceedings of the Second International Conference on
Genetic Algorithms, pages 41–49, 1987.
[12] D.E. Goldberg. Genetic Algorithms in search, Optimization and Ma-

chine Learning. Addison Wesley, 1989.
[13] P. Greistorfer and S. Voß. Controlled pool maintenance in combinato-

rial optimization. In C. Rego and B. Alidaee, editors, Adaptive Memory
and Evolution: Tabu Search and Scatter Search. Kluwer Academic Pub-
lishers, Boston, To appear.
[14] A. Hertz and M. Widmer. Guidelines for the use of meta-heuristics in

combinatorial optimization. European Journal of Operational Research,
151(2):247–252, 2003.
[15] J.H. Holland. Adaptation in natural and artificial systems. Technical

report, University of Michigan, Ann Arbor, 1975.
[16] M. Laguna, R. Martı́, and V. Campos. Intensification and diversifica-

tion with elite tabu search solutions for the linear ordering problem.
Computers and Operations Research, 26:1217–1230, 1999.
[17] E.L. Lawler. A ’pseudo-polynomial’ algorithm for sequencing jobs to

minimize total tardiness. Annals of Discrete Mathematics, 1:331–342,
1977.
[18] S.W. Mahfoud. Crowding and preselection revisited. In R. Manner and

B. Manderick, editors, Parallel Problem Solving from Nature, pages 27–
36, Elsevier, Amsterdam, 1992.
[19] M. Mauldin. Maintaining diversity in genetic search. In Proceedings of

the National Conference on Artificial Intelligence, pages 247–250, 1984.
[20] P. Moscato. On evolution, search, optimization, genetic algorithms and

martial arts: Towards memetic algorithms. Technical Report C3P 826,
Caltech Concurrent Computation Program, 1989.
15
[21] M-C. Portmann. Genetic algorithm and scheduling: a state of the art
and some propositions. In Proceedings of the workshop on production
planning and control, pages I–XIV, Mons, Belgium, 1996.
[22] C. Prins, M. Sevaux, and K. Srensen. A genetic algorithm with pop-

ulation management for the CARP. In Proceedings of TRISTAN V,
Guadaloupe, 2004.
[23] C.R. Reeves. Using genetic algorithms with small populations. In

S. Forrest, editor, Proceedings of the 5th International Conference on
Genetic Algorithms, pages 92–99, Morgan Kaufman, San Mateo, 1993.
Morgan Kaufmann.
[24] C.R. Reeves. Genetic algorithms for the operations researcher. IN-
FORMS Journal on Computing, 9:231–250, 1997.
[25] S. Ronald. Distance functions for order-based encodings. In D. Fogel,

editor, Proceedings of the IEEE Conference on Evolutionary Computa-
tion, pages 641–646, IEEE Press, New York, 1997.
[26] S. Ronald. More distance functions for order-based encodings. In Pro-

ceedings of the IEEE Conference on Evolutionary Computation, pages
558–563, IEEE Press, New York, 1998.
[27] K. Sörensen. Distance measures based on the edit distance for

permutation-type representations. In A. Barry, editor, Proceedings of
the Workshop on Analysis and Design of Representations and Operators
(ADoRo), GECCO Conference, pages 15–21, Chicago, 2003.
[28] K. Sörensen. A framework for robust and flexible optimisation using

metaheuristics with applications in supply chain design. PhD thesis,
University of Antwerp, 2003.
[29] E. Ukkonen. Finding approximate patterns in strings. Journal of Al-

gorithms, 6:132–137, 1985.
[30] A.H.C. van Kampen and L.M.C. Buydens. The ineffectiveness of recom-
bination in a genetic algorithm for the structure elucidation of a hep-
tapetide in torsion angle space. A comparison to simulated annealing.
Chemometrics and Intelligent Laboratory Systems, 36:141–152, 1997.
[31] R.A. Wagner and M.J. Fischer. The string-to-string correction problem.
Journal of the Association for Computing Machinery, 21:168–173, 1974.
16
A genetic algorithm for a
bi-objective capacitated arc routing problem
P. Lacomme1 , C. Prins2 and M. Sevaux3∗
1 Université Blaise Pascal, LIMOS, Campus Universitaire des Cézeaux, BP 10125, F-63177 Aubière Cedex, France
2 Université de Technologie de Troyes, LOSI, 12 Rue Marie Curie, BP 2060, F-10010 Troyes Cedex, France
3 Université de Valenciennes, LAMIH/SP, Le Mont Houy, F-59313 Valenciennes Cedex 9, France
June 2003
Abstract
The Capacitated Arc Routing Problem (CARP) is a very hard vehicle routing problem
for which the objective – in its classical form – is the minimisation of the total cost of
the routes. In addition, one can seek to minimize also the cost of the longest trip. In
this paper, a multi-objective genetic algorithm is presented for this more realistic CARP.
Inspired by the second version of the Non-dominated Sorted Genetic Algorithm frame-
work (NSGA-II), the procedure is improved by using good constructive heuristics to seed
the initial population and by including a local search procedure. The new framework and
its different flavours are appraised on three sets of classical CARP instances comprising
81 files. Yet designed for a bi-objective problem, the best versions are competitive with
state-of-the-art metaheuristics for the single objective CARP, both in terms of solution
quality and computational efficiency: indeed, they retrieve a majority of proven optima
and improve two best-known solutions.
Keywords: Capacitated Arc Routing Problem; Multi-objective optimization; Meta-

heuristic; Genetic algorithm.
1 Introduction: the CARP and its bi-objective version

The Capacitated Arc Routing Problem (CARP) in its undirected version is defined as follows.
Let G = (V, E) be an undirected graph with a set V of n nodes and a set E of m edges. A
fleet of identical vehicles of capacity W is based at a depot node s. The number of vehicles
is either fixed or left as a decision variable. A subset of τ edges, denoted as required edges
or tasks, require service by a vehicle. All edges can be traversed any number of times. Each
edge (i, j) has a traversal cost cij and a demand qij . All costs and demands are integers. The
goal is to determine a set of vehicle trips (routes) of minimum total cost, such that each trip
starts and ends at the depot, each required edge is serviced by one single trip, and the total
demand handled by any vehicle does not exceed W .
∗
Corresponding author. Tel: +33-327-51-1326, fax: +33-327-51-1310.
E-mail address: Marc.Sevaux@univ-valenciennes.fr (M. Sevaux)
1
Since the CARP is N P-hard, large scale instances must be solved in practice with heuris-
tics. Among fast constructive methods, one can cite for instance Path-Scanning [1], Augment-
Merge [2] and Ulusoy’s splitting technique [3]. Available metaheuristics include tabu search
algorithms [4,5], a tabu search combined with scatter search components [6], a variable neigh-
borhood search [7], a guided local search [8], and hybrid genetic algorithms [9, 10]. All these
heuristics can be evaluated thanks to tight lower bounds [11, 12].
CARP problems are raised by operations on street networks, e.g. urban waste collection,
snow plowing, sweeping, gritting, etc. Economically speaking, the most important application
certainly is municipal refuse collection. In that case, nodes in G correspond to crossroads,
while edges in E model street segments that connect crossroads. The demands are amounts
of waste to be collected. The costs correspond either to distances or durations.
The single objective CARP only deals with minimizing the total cost of the trips. In fact,
most waste management companies are also interested in balancing the trips. For instance,
in Troyes (France), all trucks leave the depot at 6 am and waste collection must be completed
as soon as possible to assign the crews to other tasks, e.g. sorting the waste at a recycling
facility. Hence, the company wishes to solve a bi-objective version of the CARP, in which
both the total duration of the trips and the duration of the longest trip (corresponding to
the makespan in scheduling theory) are to be minimized. A multiobjective genetic algorithm
(MOGA) for this bi-objective CARP is investigated in this paper which extends preliminary
results presented at the EMO 2003 conference [13]. As the two objectives are conflicting, the
fleet size (equal to the number of trips) is not imposed, in order to get smaller makespan
values. This is not a problem because fleet size is naturally bounded above by τ (when each
trip contains one single task).
For a complete introduction to multi-objective optimisation, we refer the reader to a recent
annotated bibliography [18] which provides a suitable entry point for general definitions and
good references. Concerning the MOGAs, two books are entirely devoted to these algorithms
[19, 20].
In the last decade, there has been a growing interest in multi-objective vehicle routing,
but all published papers deal with node routing problems, e.g. [14–17]. Moreover, apart from
the total cost of the trips, they only consider additional objectives such as minimizing time
window violations or balancing vehicle loads, but not makespan. To the best of our knowledge,
this paper presents the first study devoted to a multi-objective arc routing problem and with
one objective related to the makespan. It is organized as follows. Section 2 recalls a general
framework of multi-objective genetic algorithm, called NSGA-II, and describes how to adapt it
for the bi-objective CARP. The inclusion of a local search procedure is discussed in Section 3.
Numerical experiments are conducted in Section 4. Section 5 brings some concluding remarks.
2 A genetic algorithm for the bi-objective CARP

Today, several MOGA frameworks are available in literature and selecting the best one for a
given MOOP is not obvious. A recent survey [21] and two comparative studies [22, 23] try to
provide guidelines for selection but these papers consider unconstrained problems, continuous
objective functions, and specific sets of benchmarks. It is very difficult to draw conclusions
for constrained combinatorial optimization problems.
Our choice for the bi-objective CARP finally turned to a MOGA framework called NSGA-
II (Non-Dominated Sorting GA) and designed by Deb [20,24]. The reasons are: a) its modular
2
and flexible structure, b) the possibility of upgrading a single-objective GA to NSGA-II
(the encoding of chromosomes and crossover operators can be recycled) and c) its successful
applications to a wide range of problems. The first subsection is necessary to introduce the
non-trivial framework of NSGA-II, while correcting minor bugs and typos from [20, 24]. The
adaptation to the bi-objective CARP is described in the second subsection.
2.1 The NSGA-II framework

Non-dominated sorting algorithm The NSGA-II algorithm computes successive genera-
tions on a population of solutions partitioned into non-dominated fronts. The non-dominated
set is identified and constitutes the non-dominated front of level 1 or front 1. This set is
temporarily disregarded from the population. The non-dominated set of the remaining popu-
lation gives the non-dominated front of level 2, and so on. This process called non-dominated
sorting is repeated until the partition is complete, see the left part of Figure 1 for an example.
f2 f2
R(1)
R(k−1)
rank 4 R(k)
R(k+1)
rank 3
R(nr)
rank 2
rank 1
f1 f1
Figure 1: Non-dominated sorting (left) and crowding distances (right)
The method is detailed in Algorithm 1. The population is implemented as a table pop of

ns solutions and pop(i) denotes the ith solution. The partition is a vector front of ns lists of
solutions indexes, i.e. i ∈ front(k) means that solution pop(i) is currently assigned to the front
of level k. Dominance between two solutions x and y is computed by the dom(x, y) boolean
function. nb better (i) counts the solutions which dominate solution pop(i) and set worse(i)
is the set of solutions dominated by pop(i).
The non-dominated sorting algorithm returns the number of fronts nf, the partition
front(1), front(2), . . . , front(nf ) and a vector rank in which rank (i) denotes the index of the
front that stores solution pop(i). A first phase computes in O(nc ·ns 2 ) nb better and set worse
values, a second one in O(ns 2 ) assigns solutions to their respective fronts and computes nf
and rank. The overall complexity is then O(nc · ns 2 ).
Crowded tournament operator In NSGA-II, each solution is assigned a fitness equal to

its non-domination level (1 being the best level). Thus, minimization of the fitness is assumed.
Each generation generates new solutions by selecting parent-solutions and applying crossover
and mutation operators. Each parent is selected using binary tournament: two solutions are
randomly selected from the population and the fittest one (with the smallest rank) is kept.
This classical selection technique is modified as follows in NSGA-II: when two solutions
belong to the same front, the tournament prefers the most isolated one, using a crowding
distance described below. The crowding distance margin(i) of a solution pop(i) is a kind of
3
Algorithm 1 procedure non dominated sort(pop,ns,front,nf,rank)
1: front(1) ← ∅
2: for i ← 1 to ns do
3: nb better(i) ← 0
4: set worse(i) ← ∅
5: for j ← 1 to ns do
6: if dom(pop(i),pop(j)) then
7: add j to set worse(i)
8: else if dom(pop(j),pop(i)) then
9: nb better(i) ← nb better(i)+1
10: end if
11: end for
12: if nb better(i) = 0 then
13: add i to front(1)
14: end if
15: end for
16: nf ← 1
17: loop
18: for all i in front(nf) do
19: rank(i) ← nf
20: for all j in set worse(i) do
21: nb better(j) ← nb better(j)-1
22: if nb better(j) = 0 then
23: add j to front(nf+1)
24: end if
25: end for
26: end for
27: exit when front(nf+1) = ∅
28: nf ← nf+1
29: end loop
measure of the search space around it which is not occupied by any other solution in the
population. The winner of a tournament between two solutions pop(i) and pop(j) can then
be computed thanks to a crowded tournament operator, implemented as a boolean function
better (i, j) which returns true if rank (i) < rank (j) or (rank (i) = rank (j)) and (margin(i) >
margin(j)). The idea is to favor the best solutions, while keeping the fronts well scattered to
prevent clusters of solutions.
To simplify, consider nc = 2 criteria like in the bi-objective CARP and a front R of nr
solutions. Let fcmin and fcmax be the minimum and maximum values of criterion fc in R,
c = 1, 2. Sort R in increasing values of the first criterion and let R(k) be the k th solution in
the sorted front. The crowding distance of R(k) is defined as follows for 1 < k < nr :
f1 (R(k) + 1) − f1 (R(k) − 1) f2 (R(k) − 1) − f2 (R(k) + 1)

margin(R(k)) = + (1)
f1max − f1min f2max − f2min
For k = 1 or k = nr , margin(R(k)) = ∞. The goal of this convention is to favour the
4
two extreme points of the front, to try to enlarge it. The right part of Figure 1 depicts the
computation of margins. The margin of R(k) is nothing but half of the perimeter of the
rectangle indicated by dashed lines.
Generation of new solutions The computation of new solutions at each iteration of

NSGA-II can be implemented as a procedure add children, specified by Algorithm 2. The
vector rank computed by the non dominated sorting procedure and the vector of crowding
distances margin are required for the crowded tournament operator. The procedure creates
ns offsprings which are added at the end of pop, thus doubling its size. Any crossover or
mutation operator can be used, depending on the problem at hand.
Algorithm 2 procedure add children(pop,ns,rank,margin)

1: input ns ← ns
2: for ns ← input ns to 2×input ns do
3: draw two distinct parents P1, P2 using the crowded tournament operator
4: combine P1 and P2, using a crossover operator, to get one new solution S
5: mutate S with a fixed probability
6: pop(ns) ← S
7: end for
NSGA-II overall structure The general structure of NSGA-II is given in Algorithm 3. An

initial population pop of ns random solutions is built by the first pop procedure and sorted by
non-domination, using Algorithm 1. The procedure get margins computes for each solution
pop(i) its crowding distance margin(i). Then, each iteration of the main loop starts by
calling add children, to create ns children which are added at the end of pop, see Algorithm 2.
Finally, the resulting population with 2 · ns solutions is reduced to a new population newpop
by keeping the ns best solutions. To do this, fronts and margins must be updated, using
non dominated sort and get margins. Starting from the front of level 1, complete fronts are
then transferred to newpop as long as possible. The first front front(i) which could not be
accomodated fully is truncated by keeping its most widely spread solutions. This is achieved
by arranging its solutions in descending order of the crowding distance values, thanks to the
margin sort procedure, and by copying the best solutions until newpop contains exactly ns
solutions. Finally, pop receives the contents of newpop for the next iteration of the GA.
2.2 GA components for the bi-objective CARP

This section describes the required components to instanciate the NSGA-II framework for
our bi-objective CARP. The representation of solutions as chromosomes and their evaluation
come from a GA [10] which is currently one of the most effective solution methods for the
single objective CARP. Therefore, only a short description is provided here: the reader is
invited to refer to [9] or [10] for more details.
2.2.1 Chromosome representation and evaluation

A chromosome is an ordered list of the τ tasks, in which each task may appear as one of its two
directions. Implicit shortest paths are assumed between successive tasks. The chromosome
5
Algorithm 3 procedure nsga2()
1: first pop(pop,ns)
2: non dominated sort(pop,ns,front,nf,rank)
3: get margins(pop,ns,front,nf,margin)
4: repeat
5: add children(pop,ns,rank,margin)
6: non dominated sort(pop,ns,front,nf,rank)
7: get margins(pop,ns,front,nf,margin)
8: newpop ← ∅
9: i←1
10: while |newpop|+|front(i)| ≤ ns do
11: add front(i) to newpop
12: i ← i+1
13: end while
14: missing ← ns-|newpop|
15: if missing 6= 0 then
16: margin sort(front,i,margin)
17: for j ← 1 to missing do
18: add the j-th solution of front(i) to newpop
19: end for
20: end if
21: pop ← newpop
22: until stopping criterion
does not include trip delimiters and can be viewed as a giant tour for a vehicle with infinite
capacity. An O(τ 2 ) procedure Split described in [10] is used to derive a least total cost
CARP solution (subject to the given sequence), by splitting the chromosome (giant tour)
into capacity-feasible tours. This technique is based on the computation of a shortest path in
an auxiliary graph, in which each arc models one possible feasible tour that can be extracted
from the giant tour. The makespan (second objective) corresponds to the maximum duration
of the trips computed by Split.
In practice, each chromosome is stored in pop as a record with three fields: the sequence of
tasks and the values computed by Split for the total cost and the makespan. The associated
solution with its detailed trips is not stored because it is required only at the end of the GA,
to print the results. It can be extracted at that time by calling Split again.
Following GA terminology, each of our chromosomes represents the genotype of a solution,
i.e. one abstraction of a solution instead of the solution itself. The actual solution (phenotype)
is materialized by a evaluation/decoding procedure, here Split. The reader may be surprised
by the priority given by Split to the total cost over the makespan. Indeed, it is possible to
adapt Split to minimize makespan [10] or even to minimize total cost, subject to an upper
bound on the makespan. In fact, all these evaluations are pertinent because they return
a non-dominated solution among the possible interpretations of the same chromosome. An
example of non-pertinent evaluation is to scan the chromosome, starting from the first task,
and to create a new trip each time the current task does not fit vehicle capacity. Clearly, such
an evaluation “wastes” the chromosome because it is dominated by Split.
A version of Split that minimizes total cost was adopted because no good rule is available
6
to choose amongst pertinent evaluations. Giving priority in that way to the total cost has
no negative impact on the overall performance of the GA. Indeed, the aim of a decoding
procedure in general is to define a mapping from the set of chromosomes into the set of
solutions and to assign reproducible values to the two objectives, but not to optimize in some
way. Optimizing is the role of the search mechanism of the GA, with its selections that favour
the fittest parents for reproduction.
2.2.2 Initial population and clone management

Most published MOGAs start from an initial population pop made of ns random chromosomes,
but including a few good solutions always accelerates the convergence. The procedure first pop
in Algorithm 3 uses three such solutions computed by classical CARP heuristics, namely
Path-Scanning [1], Augment-Merge [2] and Ulusoy’s heuristic [3]. Each heuristic solution is
converted into a chromosome by concatenating its trips. The chromosome is then re-evaluated
by Split because the resulting solution sometimes dominates the heuristic solution.
Although this is not part of the NSGA-II framework, a sufficient condition is used to
prevent identical solutions (clones): in the initial population and in the subsequent iterations
of the GA, there can be at most one solution with two given values of the objectives, i.e. two
solutions cannot occupy the same point in the graphical representation of the fronts (remember
that all costs are integers). This technique was included to avoid a progressive colonization
of the population by clones (typically 20 to 30% after 100 iterations).
2.2.3 Crossover operators

The advantage of chromosomes without trip delimiters is to allow the use of classical crossovers
designed for permutation chromosomes, like the Linear Order Crossover (LOX) and the Order
Crossover (OX). Like the single-objective CARP studied in [10], the OX crossover gave the
best results after some preliminary testing. This crossover must be slightly adapted because
the chromosomes are not exactly permutations: each required edge may appear as one of its
two directions. Implementation details can be found in [10].
Hence, the crossover used in the add children procedure of Algorithm 2 is in fact OX.
One of the two children is selected at random and is evaluated by Split. It is added to pop
only if no solution exists with the same objective values, else it is discarded. Because of such
rejections, more than ns crossovers may be required to double the size of pop: the FOR loop
of Algorithm 2 must be replaced by a REPEAT . . . UNTIL ns = 2 × input ns.
2.2.4 Stopping conditions

There is no standard technique to define reliable stopping conditions for stopping a MOGA.
In combinatorial optimization and in the single-objective case, metaheuristics can be stopped
after a fixed number of iterations without improving the current best solution, or by using a
variance criterion. The ways of extending these criteria to several objectives still raise endless
discussions in the MOO community. The underlying problem, discussed in the numerical
experiments of Section 4, is to define a pertinent comparison between the final sets of efficient
solutions computed by two different algorithms on the same instance. For this study, the
MOGA is simply stopped after a fixed number of iterations. More sophisticated stopping
criteria had no noticeable effect on the final results.
7
3 Local search procedures
In single-objective optimization, it is well known that a standard GA must be hybridized
with a local search procedure to be able to compete with state-of-the-art metaheuristics like
tabu search. Such GAs, called memetic algorithms by Moscato [25], have raised a growing
interest in the last decade: for instance, a dedicated conference (WOMA) is now entirely
devoted to them. Hybridization is still rarely used in multi-objective optimization, and only
a few authors like Jaszkiewicz [26] start investigating this area. Several difficulties are raised,
for instance: a) the notion of local search must be clarified when several objectives must be
optimized, b) several locations are possible in the existing MOGA frameworks for calling a
local search, and c) the local search must not disturb the search mechanism of the MOGA, for
instance by concentrating the solutions into small clusters. This section describes the moves
and the general structure of our local search procedure, the criteria used to accept a move,
and the possible ways of integrating local search in the MOGA.
3.1 Moves tested and general structure

Our local search procedures work on the individual routes, not on the sequence of tasks defined
by a chromosome. Hence, the input chromosome is first converted into a set of routes. The
moves include the removal of one or two consecutive tasks from a route, with reinsertion at
another position, and 2-opt moves. All moves may involve one or two routes and the two
traversal directions of an edge are tested in the reinsertions. Each iteration of the local search
scans all these moves in O(τ 2 ) to find the first improving move (the criteria for deciding if
there is an improvement are discussed in the next subsection). The whole local search stops
when no more improvement can be found. The trips are then concatenated into a chromosome
(giant tour), which is reevaluated by Split.
3.2 Acceptance criteria for a move

Four acceptance criteria accept(S, S 0 ) were tested to accept a move that transforms the in-
cumbent solution S into a neighbor solution S 0 . This has lead to four local search procedures
called LS1 to LS4. In what follows, f1 (S) and f2 (S) respectively denote the total cost and
the makespan of solution S and w1 is a real number in [0, 1].
LS1: accept(S, S 0 ) = f1 (S 0 ) − f1 (S) < 0;
LS2: accept(S, S 0 ) = f2 (S 0 ) − f2 (S) < 0;
LS3: accept(S, S 0 ) = ((f1 (S 0 ) − f1 (S) ≤ 0) and (f2 (S 0 ) − f2 (S) < 0))

or ((f1 (S 0 ) − f1 (S) < 0) and (f2 (S 0 ) − f2 (S) ≤ 0));
LS4: accept(S, S 0 ) = w1 · (f1 (S 0 ) − f1 (S)) + (1 − w1 ) · (f2 (S 0 ) − f2 (S) < 0.
LS1 performs its descent on the total cost, LS2 on the makespan, while LS3 accepts the
new solution if and only if it dominates the current one. The last local search is a weighted
sum method which scalarizes the two objectives into a single objective, by multiplying each
objective by a user-supplied weight. In these four versions, no normalization of the objectives
is required because the total duration of the trips and the makespan are commensurable and
of the same order of magnitude.
8
3.3 Location of local search in the MOGA
The easiest way to include a local search procedure in the MOGA probably is to replace the
mutation in Algorithm 2. The mutation can be suppressed because diversification is inherent
to the NSGA-II framework, thanks to crowding distances. However, a systematic local search
plays against a good dispersal of solutions in the population. This is why, in practice, an
offspring undergoes a non-systematic local search with a fixed probability p LS , typically 10%.
This location of the local search was the only one tested for LS1, LS2 and LS3. It is called
local search on children in the numerical experiments of Section 4.
In LS4, the weight w1 that determines the descent direction must be computed. A random
selection as suggested in [27] gives poor results. Murata et al. [28] have designed a MOGA
(not based on NSGA-II) for a pattern classification problem, in which a local search like LS4
is applied to all solutions of front 1, with a descent direction which depends on the location
of solutions on this front, see Figure 2. The aim is to improve the front, with emphasis on the
two extreme solutions, while preserving the spacing between solutions. To compute w 1 for a
given solution S, these authors use Equation 2 in which fcmin and fcmax respectively denote
the minimum and maximum values of criterion fc , c = 1, 2.
Ã ! ,Ã !
f1 (S) − f1min f1 (S) − f1min f2 (S) − f2min
w1 = + max (2)
f1max − f1min f1max − f1min f2 − f2min
f2
rank 3
rank 2
rank 1
f1
Figure 2: Variable descent directions in Murata’s MOGA
For the instances tested in Section 4 for the bi-objective CARP, the first front contains on
average less than 1/10 of the solutions and applying LS4 to the first front only is not agressive
enough. However, by computing f1min , f1max , f2min and f2max on the whole population instead
of the first front, LS4 can be applied to any solution and with a descent direction that depends
on its position in pop. This option was finally selected for our MOGA.
Two possible locations can be used for LS4. As for LS1, LS2 and LS3, it can be called in
add children with a given probability pLS , in lieu of mutation (local search on children). In
that case, f1min , f1max , f2min and f2max are computed on the whole population, at the beginning
of the procedure, and the descent direction for child S is computed using Equation 2.
The other location consists in applying periodically LS4 to a fraction of the fronts, just
after the non-dominated sort in Algorithm 3. Since this local search phase may change the
fronts, it must be followed by a second call to non dominated sort. To keep approximately an
overall local search rate of 10%, LS4 is applied every k iterations of the MOGA to k × ns/10
9
solutions of pop, for instance 10% of solutions every iteration or 20% every two iterations. In
fact, better results were obtained by applying LS4 to all solutions every k = 10 iterations,
probably because of a better oscillation between diversification and intensification. We call
this way of including LS4 periodic local search.
4 Computational experiments
All developpements were done in Delphi 7 (Borland°r
) and tested on a Pentium IV laptop
computer clocked at 1.8 GHz under Windows 2000. The following subsections describe the
instances used, the MOGA versions selected, the evaluation criteria, the results, and two
graphical examples.
4.1 Instances tested

The computational evaluation is based on three sets of standard CARP instances which are
used by most authors working on the single-objective CARP [4–12]. The first set corresponds
to the 23 gdb files from Golden, DeArmon and Baker [1], with 7 to 27 nodes and 11 to 55
edges, all required. The original set contains 25 files but two instances (gdb8 and gdb9) are
inconsistent and never used. The second set gathers the 34 val files proposed by Belenguer
and Benavent in [12], which have 24 to 50 nodes and 34 to 97 edges, all required. The last
set provides the 24 egl instances generated by Belenguer and Benavent [12], by perturbing
the demands in three rural networks supplied by Eglese. These egl files are bigger (77 to 140
nodes, 98 to 190 edges) and in some instances not all edges are required. All these files can
be downloaded from the internet, at address http://www.uv.es/~belengue/carp.html.
4.2 Test protocol and parameters

A preliminary testing phase was required to fix the population size, the crossover (OX or
LOX), the number of iterations and to evaluate the impact of good heuristic solutions in the
initial population. This has lead to the following features which are shared by all MOGA
versions in the sequel: a population of ns = 60 solutions with three good initial solutions, the
OX crossover and 100 to 200 iterations corresponding to 6,000 to 12,000 crossovers. Path-
Scanning [1], Augment-Merge [2] and Ulusoy’s heuristic [3] were selected to provide good
initial solutions. They can be discarded without changing the quality of final solutions, but
in that case the MOGA converges more slowly, in 300 to 400 iterations.
A basic version of the MOGA without local search, called MO1, was selected as a reference
algorithm. Eight other versions listed in Table 1 were prepared for comparison. They differ
from MO1 by the number of iterations and the kind of local search.
4.3 Evaluation criteria

Comparing two different algorithms for solving the same problem is a central issue in multi-
objective optimization, especially when the final choices of the decision maker are unknown.
Several criteria were published, either to evaluate the quality of front 1 at the end of one
given algorithm, or to provide a relative comparison between the final fronts computed by
two algorithms.
10
Table 1: MOGA versions tested
Version Iterations LS type LS location
MO1 100 None Irrelevant
MO2 100 1 On children, rate 10%
MO8 100 4 Periodic, every 10 iterations
MO9 200 4 Periodic, every 10 iterations
This study uses a simple measure proposed by Riise [29] to compare two fronts. Given
a front F to be evaluated and a reference front R (defined by MO1 in our case), a kind of
distance is measured between each solution of F and its projection onto the extrapolated
reference front. This extrapolated front is a piecewise continuous line (in the bi-objective
case) going through each solution and extended beyond the two extreme solutions, like in
Figure 3. These distances are multiplied by −1 when the solution to compare is below the
extrapolated front, like d4 in the figure. The proposed measure µ(F, R) is the sum of the
signed distances between the solutions of F and their respective projections on the reference
front, e.g. µ = d1 + d2 + d3 − d4 in the figure.
f2
reference front
d1
front to compare
d2
d3
d4
extrapolated
front
f1
Figure 3: Examples of distances to the extrapolated front
Hence, for one given instance, a version of the MOGA with a final front F outperforms
MO1 with its final front R if µ(F, R) < 0. Since µ depends on the number of solutions in F , a
normalized measure µ̄(F, R) = µ(F, R)/|F | will be also used. The measure µ is a bit arbitrary
and of course subject to criticism. For example, how can we measure the difference between
two fronts, when one is well spread while the other displays several gaps between clusters of
solutions? Nevertheless, its merit is to exist and to help us rank the tested protocols. Because
of no a-priori knowledge of the solutions which lie on the efficient front, such measures should
be completed by a visual inspection.
Very recently, some authors coming from the combinatorial optimization community like
11
Jaszkiewicz [27] have claimed that the authors of multi-objective metaheuristics should prove
to be competitive to single objective metaheuristics in terms of quality of solutions and
computational efficiency. Since several metaheuristics are available for the single objective
CARP [4–10], they can be used to evaluate the total cost of the leftmost solution of front 1,
thus completing the measures µ and µ̄. Two metaheuristics were selected for that purpose in
our computational evaluation: CARPET, the most efficient tabu search algorithm, designed
by Hertz et al. [5], and the memetic algorithm (MA) proposed by Lacomme et al. [10].
Our tables of results compare also the leftmost and rightmost solutions with two lower
bounds. The lower bound LB1 for the total cost is the one proposed by Belenguer and
Benavent [12]. Concerning makespan, a trivial lower bound LB2 corresponds to the longest
trip, when each trip performs one single task: LB2 = max{Dsi + cij + Djs | [i, j] ∈ E}, where
s is the depot node and Dij the cost of a shortest path from node i to node j in G.
4.4 Results
4.4.1 Format of tables
For each set of benchmarks, the next subsections present one synthetic table with the results
obtained by CARPET, the MA and each MOGA version, see Table 2 for instance. This table
indicates for each algorithm the following average values: the number of solutions in the final
set of non-dominated solutions F , the deviation Dev LB1 of the leftmost solution to the lower
bound on the total cost LB1 , the number of times LB1 hits this bound is reached (number
of proven optima), the deviation Dev LB2 of the rightmost solution to the makespan lower
bound LB2 , the number of proven optimal makespans LB2 hits, the measure µ for a reference
front provided by MO1, the normalized measure µ̄ and the CPU time in seconds on a 1.8
GHz Pentium-IV PC. For CARPET and the MA, the only applicable columns are Dev LB 1 ,
LB1 hits and the running time. These fields are in boldface when a MOGA is at least as
good as a single objective metaheuristic.
For each synthetic table, the results of the best MOGA version in terms of µ̄ are listed
instance by instance in a detailed table given in the Appendix, see Table 5 for instance. The
three first columns indicate the instance name, the number of nodes n and the number of
edges m. The number of tasks τ is cited only for egl files, because it is always equal to the
number of edges in the other benchmarks. These columns are followed by the total cost and
the running time achieved by CARPET (listed in [5], except the egl files whose solution values
are given in [12], but without running times) and by the memetic algorithm [10].
In order to have a fair comparison with the MOGAs, the solution values are given for the
standard setting of parameters of these metaheuristics and the running times are scaled for
the 1.8 GHz PC used in this study. The MA in [10] was executed on a 1 GHz Pentium-III
PC and its running times are here multiplied by 0.555. CARPET was tested on a 195 MHz
Silicon Graphics Indigo-2 workstation which is 5.43 times slower than the Pentium-III. Hence,
the execution times provided by Hertz et al. are multiplied by 0.555/5.43 = 0.1022 in our
study. The best-known solution values (total cost) found by CARPET and the MA with
various settings of parameters are given in the BKS column.
The eight remaining columns concern the MOGA: the number of efficient solutions |F |,
the total costs and makespans for the leftmost and rightmost solutions (f 1lef t , f2lef t , f1right
and f2right ), the running time in seconds and the measures µ and µ̄. Each detailed table ends
with three lines that provide for each column its average value (deviation to a bound, running
12
time or measure), the worst value and the number of proven optima.
4.4.2 Results for gdb files

They are given in Table 2. All versions with a local search (MO2 to MO9) are much better
than MO1. As expected, MO2 with its local search LS1 on the total cost yields a good
leftmost point but a poor rightmost point, while the contrary holds for MO3 with its LS2
on makespan. MO4 with its local search LS3 (accepting a move if the result dominates the
incumbent solution) is almost as good as CARPET but performing 200 iterations instead
of 100 (MO5) brings no improvement. The versions which periodically apply the directional
local search LS4 to all solutions (MO8 and 9) respectively outperform MO6 and MO7 in which
LS4 is called on children. The best versions for 100 and 200 iterations are respectively MO8
and MO9, they have little effect on the two measures but improve the two extreme solutions.
They outperform CARPET for the average deviation to LB1 but do not find more optima
and are a bit slower.
Table 2: Average results for the 23 gdb files

Method |F | Dev LB1 LB1 hits Dev LB2 LB2 hits µ µ̄ Time (s)
CARPET – 0.48 18 – – – – 5.01
MA – 0.15 21 – – – – 2.94
MO1 3.87 3.81 5 29.64 9 0.00 0.00 1.94
MO2 3.52 0.47 17 30.43 7 -18.15 -4.61 6.01
MO3 3.78 3.40 5 24.97 11 -5.56 -1.74 4.56
MO4 3.26 0.56 16 22.07 11 -22.82 -6.48 6.53
MO5 3.30 0.56 16 20.28 11 -23.64 -6.45 11.53
MO6 3.39 0.56 16 22.00 10 -25.14 -6.39 6.46
MO7 3.43 0.48 16 21.33 11 -24.07 -6.35 12.94
MO8 3.39 0.39 18 21.18 12 -23.94 -6.45 7.17
MO9 3.43 0.36 18 20.95 12 -23.49 -6.53 13.79
The results of the best version MO9 are detailed in Table 5, in the appendix. They show
that MO9 is very robust, since its worst deviation to LB1 is 2.23% vs 4.62% for CARPET.
No total cost obtained by MO9 is improved by the other versions. In particular, the two
instances gdb10 and gdb14 remain open.
4.4.3 Results for val files

The synthetic results of Table 3 show the same hierarchy as for gdb files: the MOGAs with LS4
outperform the MOGAs with LS3, which in turn outperform the versions with single objective
local searches and the version without local search. But, this time, LS4 gives better results
when it is called on children (MO6-MO7). The convergence seems slower, since noticeable
improvements can be found by allowing 200 iterations. The average running time strongly
increases but is still reasonable (1.3 minutes maximum). It is mainly consumed by the calls
to local searches. All MOGAs with LS3 or LS4 (MO4 to MO9) outperform CARPET for the
average deviation to LB1 (MO4 is even a bit faster) but 200 iterations are required to find
more optima.
Roughly speaking, the best version MO7 halves the average deviation to LB 1 obtained by
CARPET, at the expense of a double running time. Its results are detailed for each instance
in Table 6, in the appendix. In addition to the 18 optimal total costs listed in this table, the
13
Table 3: Average results for the 34 val files
Method |F | Dev LB1 LB1 hits Dev LB2 LB2 hits µ µ̄ Time (s)
CARPET – 1.90 15 – – – – 35.48
MA – 0.61 22 – – – – 21.31
MO1 5.91 8.52 1 37.73 5 0.00 0.00 3.21
MO2 4.76 1.23 17 32.06 5 -85.43 -16.18 27.80
MO3 6.06 7.99 3 30.90 7 -7.33 -2.31 11.74
MO4 4.88 1.36 15 21.65 8 -109.05 -20.87 32.50
MO5 5.29 1.01 19 20.59 9 -125.98 -21.37 65.20
MO6 5.00 1.16 15 21.50 10 -113.07 -21.30 38.20
MO7 5.24 0.99 18 19.62 10 -121.70 -21.85 76.17
MO8 5.09 1.33 16 20.19 10 -114.11 -20.61 44.53
MO9 5.32 1.13 17 19.43 10 -124.03 -21.51 83.06
other versions have found optima for the following instances: val2b (259), val2c (457) and
val10c (446). Moreover, they improve MO7 for val4d (536 vs 539), val5d (593 vs 595) and
val9d (397 vs 399). Unfortunately, no open instance is broken and best-known solutions are
not improved.
4.4.4 Results for egl files

The lower bound LB1 proposed by Belenguer and Benavent is never tight on the egl files.
This is why the number of optimal total costs is replaced in Table 4 by the number of best-
known solutions retrieved or (in brackets) improved (column BKS ). All MOGAs except MO1
and MO3 outperform CARPET in terms of average deviation to the lower bound on the total
cost and retrieve some best-known solutions (CARPET: 0). MO9 is even able to improve
two best-known solutions, namely egl-e3-c (10349 vs 10369) and egl-s3-b (14004 vs 14028).
Contrary to gdb and val files, the MOGAs reach the optimal makespan for a majority of
instances and become faster than the memetic algorithm. In particular, the last MOGA finds
all optimal makespans except one. Average running times are still acceptable: less than 5
minutes. For the detailed results in appendix (Table 7), we preferred to select MO9 with
the lowest deviation to LB1 and its two improved instances, rather than MO5 which has a
slightly better measure µ̄.
Table 4: Average results for the 24 egl files

Method |F | Dev LB1 BKS Dev LB2 LB2 hits µ µ̄ Time (s)
CARPET – 4.74 0 – – – – Unknown
MA – 2.47 19 – – – – 292.77
MOGA1 5.83 13.24 0(0) 1.87 15 0.00 0.00 2.77
MOGA2 5.29 4.24 1(0) 0.94 16 -1234.01 -256.65 158.61
MOGA3 6.67 13.32 0(0) 0.17 20 -63.93 -15.98 35.80
MOGA4 5.21 4.47 0(0) 0.27 19 -1333.52 -335.11 148.50
MOGA5 6.00 3.84 1(0) 0.10 21 -1586.62 -360.97 263.49
MOGA6 5.33 4.47 1(0) 0.08 22 -1435.73 -324.27 153.48
MOGA7 5.83 4.00 1(0) 0.07 22 -1573.80 -302.97 290.55
MOGA8 5.96 4.11 1(0) 0.06 22 -1592.01 -344.66 159.01
MOGA9 5.79 3.69 2(2) 0.05 23 -1531.94 -350.76 268.99
BKS gives the number of best-known solutions retrieved or (in brackets) improved by the MOGA.
14
4.5 Graphical examples
Graphical representations are well suited to visualize the behavior of bi-objective optimization
methods. The kind of progression performed by our MOGAs is illustrated in Figure 4. The
graphic represents in the objective space the random initial solutions (symbol +), the three
good heuristic solutions (symbol ×) and the final population (symbol × +) computed by MO9
on the egl-e3-a instance (77 nodes, 98 edges and 87 tasks). Remember that MO9 is the version
which performs 200 iterations and applies LS4 to all solutions, every 10 iterations.
4000
Initial Population
Initial heuristics
Final population
3500
3000
2500
Makespan
2000
1500
1000
500
6000 8000 10000 12000 14000 16000 18000 20000
Total cost
Figure 4: Convergence of MO9 on instance egl-e3-a
One can notice that the initial heuristic solutions are already quite close to the final
population. As already mentioned, adding such solutions is very useful to accelerate the
search. Figure 5 provides a magnified representation of the final population. A dashed line
is used to emphasize efficient solutions.
The impact of the other key-component, the local search procedure, can be illustrated
on the same instance by comparing the efficient set computed by the basic version MO1 (no
local search, 100 iterations) with the one obtained by the best version MO9 (periodic local
search LS4, 200 iterations). See Figure 5. Even if the total number of iterations is increased
for MO1, the results are only slightly improved. As for single objective GAs, hybridization
with local search is necessary to obtain efficient MOGAs. Here for instance, MO9 decreases
the best total cost by 10.55% (from 6659 to 5956) and the best makespan by 10.97% (from
921 to 820). Moreover, more efficient solutions are obtained and they are better spread, thus
providing the decision maker with a wider choice.
5 Concluding remarks
In industry, most decision makers wish to take several criteria into account and the bi-objective
CARP addressed in this article is a typical example. The combinatorial optimization commu-
nity can help research in multi-objective optimization, by bringing a rich toolbox, very efficient
algorithms, and a stricter methodology to evaluate and compare different algorithms.
15
No Local search (MO1) - Efficient solutions
Local search (MO9) - Efficient solutions
1000 Local search (MO9) - Final population
950
Makespan
900
850
800
5800 6000 6200 6400 6600 6800 7000 7200 7400 7600 7800 8000
Total cost
Figure 5: Impact of local search for instance egl-e3-a
In this paper, the implementation of the NSGA-II multi-objective genetic algorithm alone
could not efficiently solve the bi-objective CARP. Two key-components were required to
enhance performance: using good heuristics (Path-Scanning, Augment-Merge and Ulusoy’s
heuristic) in the initial population and adding a local search able to improve solutions for
both criteria. It is worth noticing that the chromosome encoding, the crossover operator and
the two proposed improvements come from a memetic algorithm successfully developped for
the single objective CARP. The main difficulty was to select a local search adapted to the
multi-objective case and to find its best location in the MOGAs.
An intensive testing on three sets of benchmarks, with a strict comparison completing a
distance measure with lower bounds and metaheuristics for the single objective case, proves
the validity and the efficiency of the proposed approach. In particular, the best versions of our
MOGAs are still competitive with single objective metaheuristics. We even believe that any
GA or memetic algorithm for a single objective problem could be generalized in the same way
to handle several objectives, while keeping its efficiency, although this should be confirmed
with other combinatorial optimization problems.
References
[1] B.L. Golden, J.S. DeArmon, and E.K. Baker. Computational experiments with algo-
rithms for a class of routing problems. Computers and Operations Research, 10(1):47–59,
1983.
[2] B.L. Golden and R.T. Wong. Capacitated arc routing problems. Networks, 11:305–315,
1981.
[3] G. Ulusoy. The fleet size and mix problem for capacitated arc routing. European Journal
of Operational Research, 22:329–337, 1985.
16
[4] J.M. Belenguer, E. Benavent, and F. Cognata. Un metaheuristico para el problema
de rutas por arcos con capacidades. In Proceedings of the 23th national SEIO meeting,
Valencia, Spain, 1997.
[5] A. Hertz, G. Laporte, and M. Mittaz. A tabu search heuristic for the Capacitated Arc
Routing Problem. Operations Research, 48(1):129–135, 2000.
[6] P. Greistorfer. A tabu-scatter search metaheuristic for the arc routing problem. Com-
puters and Industrial Engineering, 44(2):249–266, 2003.
[7] A. Hertz and M. Mittaz. A variable neighborhood descent algorithm for the undirected
Capacitated Arc Routing Problem. Transportation Science, 35(4):425–434, 2001.
[8] P. Beullens, L. Muyldermans, D. Cattrysse, and D. Van Oudheusden. A guided local

search heuristic for the Capacitated Arc Routing Problem. European Journal Of Opera-
tional Research, 147(3):629–643, 2003.
[9] P. Lacomme, C. Prins, and W. Ramdane-Chérif. A genetic algorithm for the Capacitated
Arc Routing Problem and its extensions. In E.J.W. Boers et al., editor, Applications of
evolutionnary computing, volume 2037 of Lecture Notes in Computer Science, pages 473–
483. Springer, 2001.
[10] P. Lacomme, C. Prins, and W. Ramdane-Chérif. Competitive memetic algorithms for

arc routing problems. Technical Report LOSI-2001-01, Laboratory for Industrial Systems
Optimization, Université de Technologie de Troyes, France, 2001. To appear in Annals
of Operations Research.
[11] A. Amberg and S. Voß. A hierarchical relaxations lower bound for the Capacitated Arc
Routing Problem. In R.H. Sprague (Hrsg.), editor, Proceedings of the 35 th Annual Hawaii
International Conference on Systems Sciences, pages DTIST02:1–10, Piscataway, 2002.
IEEE.
[12] J.M. Belenguer and E. Benavent. A cutting plane algorithm for the Capacitated Arc
Routing Problem. Computers and Operations Research, 30(5):705–728, 2003.
[13] P. Lacomme, C. Prins, and M. Sevaux. Multi-objective Capacitated Arc Routing Prob-
lem. In C.M. Fonseca et al., editor, Evolutionary multi-criterion optimization (Proceed-
ings of EMO 2003, Faro, Portugal), volume 2632 of Lecture Notes in Computer Science,
pages 550–564. Springer, 2003.
[14] Y.B. Park and C.P. Koelling. An interactive computerized algorithm for multicriteria
vehicle routing problems. Computers and Industrial Engineering, 16:477–490, 1989.
[15] A. Corberan, E. Fernandez, M. Laguna, and R. Martı́. Heuristic solutions to the problem
of routing school buses with multiple objectives. Journal of the Operational Research
Society, 53(4):427–435, 2002.
[16] S.C. Hong and Y.B. Park. A heuristic for bi-objective vehicle routing problem with time
windows constraints. International Journal of Production Economics, 62:249–258, 1999.
17
[17] W. Sessomboon, K. Watanabe, T. Irohara, and K. Yoshimoto. A study on multi-objective
vehicle routing problem considering customer satisfaction with due-time (the creation of
pareto optimal solutions by hybrid genetic algorithm). Transaction of the Japan Society
of Mechanical Engineers, 1998.
[18] M. Ehrgott and X. Gandibleux. Multiobjective Combinatorial Optimization, M. Ehrgott

and X. Gandibleux (eds) , volume 52 of International Series in Operations Research and
Management Science, pages 369–444. Kluwer, 2002.
[19] C.A. Coello Coello, D.A. Van Veldhuizen, and G.B. Lamont. Evolutionary algorithms
for solving multi-objective problems. Kluwer, New York, 2002.
[20] K. Deb. Multi-objective optimization using evolutionary algorithms. Wiley, Chichester,

UK, 2001.
[21] C.A. Coello Coello. An updated survey of GA-based multiobjective optimization tech-
niques. ACM Computing Surveys, 32(2):109–143, 2000.
[22] K. Deb. Multi-objective genetic algorithms: problem difficulties and construction of test
problems. Evolutionary Computation, 7(3):205–230, 1999.
[23] E. Zitzler, K. Deb, and L. Thiele. Comparison of multiobjective evolutionary algorithms:

empirical results. Evolutionary Computation, 8(2):173–195, 2000.
[24] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan. A fast and elitist multi-objective ge-
netic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation, 6(2):182–
197, 2002.
[25] P. Moscato. New ideas in optimization D. Corne, M. Dorigo and F. Glover (eds), chapter
Memetic algorithms: a short introduction, pages 219–234. McGraw-Hill, Maidenhead,
UK, 1999.
[26] A. Jaszkiewicz. Genetic local search for multiple objective combinatorial optimization.
European Journal of Operational Research, 137(1):50–71, 2001.
[27] A. Jaszkiewicz. Do multiple objective metaheuristics deliver on their promises? a com-

putational experiment on the set covering problem. IEEE Transactions on Evolutionary
Computation, 7(2):133–143, 2003.
[28] T. Murata, H. Nozawa, H. Ishibuchi, and M. Gen. Modification of local search direc-
tions for non-dominated solutions in cellular multiobjective genetic algorithms for pattern
classification problems. In C.M. Fonseca et al., editor, Evolutionary multi-criterion op-
timization (Proceedings of EMO 2003, Faro, Portugal), volume 2632 of Lecture Notes in
Computer Science, pages 593–607. Springer, 2003.
[29] A. Riise. Comparing genetic algorithms and tabu search for multi-objective optimization.
In Abstract conference proceedings, page 29, Edinburgh, UK, July 2002. IFORS.
18
Table 5: Detailed results of MO9 for gdb files
File n m LB1 LB2 CARPET Time MA Time BKS |F | f1lef t f2lef t f1right f2right Time µ µ̄
gdb1 12 22 316 63 316* 1.75 316* 0.00 316* 3 316* 74 337 63* 8.36 -30.08 -10.03
gdb2 12 26 339 59 339* 2.87 339* 0.24 339* 3 339* 69 395 59* 10.27 -18.86 -6.29
gdb3 12 22 275 59 275* 0.04 275* 0.03 275* 4 275* 65 339 59* 8.57 -31.44 -7.86
gdb4 11 19 287 64 287* 0.05 287* 0.00 287* 3 287* 74 350 64* 7.07 -11.79 -3.93
gdb5 13 26 377 64 377* 3.10 377* 0.06 377* 6 377* 78 447 64* 10.46 -12.78 -2.13
gdb6 12 22 298 64 298* 0.47 298* 0.09 298* 4 298* 75 351 64* 8.16 -40.47 -10.12
gdb7 12 22 325 57 325* 0.00 325* 0.03 325* 3 325* 68 381 61 8.55 -14.02 -4.67
gdb10 27 46 344 38 352 33.85 350 22.10 348 4 350 44 390 38* 22.45 -19.35 -4.84
gdb11 27 51 303 37 317 29.92 303* 3.93 303* 3 309 43 333 37* 27.17 -31.77 -10.59
gdb12 12 25 275 39 275* 0.86 275* 0.03 275* 4 275* 71 297 54 10.24 -33.43 -8.36
gdb13 22 45 395 43 395* 1.27 395* 0.70 395* 5 395* 81 421 64 33.68 -105.14 -21.03
gdb14 13 23 448 93 458 11.45 458 5.43 458 4 458 97 547 93* 7.26 -64.54 -16.13
19
gdb15 10 28 536 128 544 1.34 536* 4.12 536* 1 544 128* 544 128* 9.03 -10.22 -10.22
gdb16 7 21 100 15 100* 0.27 100* 0.03 100* 3 100* 21 112 17 7.66 -1.36 -0.45
gdb17 7 21 58 8 58* 0.00 58* 0.00 58* 2 58* 15 60 13 9.66 -4.36 -2.18
gdb18 8 28 127 14 127* 0.94 127* 0.03 127* 4 127* 27 135 19 10.75 -13.78 -3.44
gdb19 8 28 91 9 91* 0.00 91* 0.03 91* 1 91* 15 91* 15 12.81 -2.77 -2.77
gdb20 9 36 164 19 164* 0.16 164* 0.06 164* 3 164* 33 178 27 18.05 -19.44 -6.48
gdb21 8 11 55 17 55* 0.11 55* 0.00 55* 2 55* 21 63 17* 4.80 -3.47 -1.73
gdb22 11 22 121 20 121* 5.27 121* 0.18 121* 5 121* 36 131 20* 8.10 -16.61 -3.32
gdb23 11 33 156 15 156* 0.63 156* 0.09 156* 4 156* 30 160 22 14.57 -10.69 -2.67
gdb24 11 44 200 12 200* 1.88 200* 1.86 200* 4 200* 26 207 20 24.03 -21.02 -5.25
gdb25 11 55 233 13 235 19.08 233* 28.41 233* 4 235 23 241 20 35.45 -22.96 -5.74
Average 0.48% 5.01 0.15% 2.94 0.13% 3.43 0.36% 47.09% 10.70% 20.95% 13.79 -23.49 -6.53
Worst 4.62% 33.85 1.78% 28.41 1.78% 6 2.23% 116.67% 23.27% 66.67% 35.45 -1.36 -0.45
Optima 18 21 21 18 1 1 12
BKS is the best solution value (total cost) found by CARPET and the MA using several settings of parameters.
CPU times in seconds on a 1.8 GHz Pentium-IV PC. Times for CARPET and MA have been scaled. See section 4.4.1 for details
Table 6: Detailed results of MO7 for val files
File n m LB1 LB2 CARPET Time MA Time BKS |F | f1lef t f2lef t f1right f2right Time µ µ̄
val1a 24 39 173 40 173* 0.02 173* 0.00 173* 1 173* 58 173* 58 20.82 -21.08 -21.08
val1b 24 39 173 40 173* 9.26 173* 8.02 173* 6 173* 61 204 42 21.39 -46.67 -7.78
val1c 24 39 235 40 245 93.20 245 28.67 245 2 245 41 248 40* 14.25 -60.72 -30.36
val2a 24 34 227 71 227* 0.17 227* 0.05 227* 6 227* 114 270 90 15.92 -39.62 -6.60
val2b 24 34 259 71 260 13.02 259* 0.22 259* 5 260 101 306 78 15.83 -26.45 -5.29
val2c 24 34 455 71 494 31.66 457 21.76 457 1 463 71* 463 71* 9.49 -7.69 -7.69
val3a 24 35 81 27 81* 0.77 81* 0.05 81* 4 81* 41 88 31 16.40 -40.05 -10.01
val3b 24 35 87 27 87* 2.79 87* 0.00 87* 4 87* 32 105 27* 14.88 -23.83 -5.96
val3c 24 35 137 27 138 41.66 138 28.23 138 1 138 27* 138 27* 10.39 -14.00 -14.00
val4a 41 69 400 80 400* 28.32 400* 0.72 400* 4 400* 134 446 92 89.86 -187.06 -46.77
val4b 41 69 412 80 416 75.66 412* 1.21 412* 9 412* 105 468 83 79.25 -402.04 -44.67
val4c 41 69 428 80 453 70.06 428* 19.11 428* 10 430 99 482 80* 73.68 -474.05 -47.40
val4d 41 69 520 80 556 233.56 541 103.26 530 1 539 80* 539 80* 50.42 -24.43 -24.43
val5a 34 65 423 72 423* 3.80 423* 1.86 423* 4 423* 141 474 96 72.75 -143.50 -35.88
val5b 34 65 446 72 448 41.40 446* 1.04 446* 8 446* 112 506 86 73.83 -151.41 -18.93
val5c 34 65 469 72 476 53.27 474 101.01 474 9 474 96 541 80 67.82 -189.67 -21.07
val5d 34 65 571 72 607 224.11 581 90.74 581 4 595 81 686 72* 51.83 -44.57 -11.14
val6a 31 50 223 45 223* 3.89 223* 0.17 223* 5 223* 75 259 56 37.08 -33.93 -6.79
20
val6b 31 50 231 45 241 26.94 233 67.34 233 7 233 68 263 50 34.93 -50.39 -7.20
val6c 31 50 311 45 329 85.18 317 52.23 317 5 317 55 329 45* 21.88 6.48 1.30
val7a 40 66 279 39 279* 6.59 279* 1.97 279* 3 279* 85 289 59 72.00 -44.27 -14.76
val7b 40 66 283 39 283* 0.02 283* 0.44 283* 4 283* 58 299 51 79.89 -81.59 -20.40
val7c 40 66 333 39 343 121.44 334 101.17 334 5 335 50 352 40 61.00 -9.03 -1.81
val8a 30 63 386 67 386* 3.84 386* 0.66 386* 7 386* 129 429 87 70.02 -172.40 -24.63
val8b 30 63 395 67 401 81.46 395* 9.95 395* 9 395* 100 455 79 66.16 -227.49 -25.28
val8c 30 63 517 67 533 147.40 527 71.46 527 7 545 74 610 67* 43.29 -180.52 -25.79
val9a 50 92 323 44 323* 28.51 323* 18.29 323* 3 326 82 333 68 171.97 -109.49 -36.50
val9b 50 92 326 44 329 59.89 326* 29.39 326* 3 326* 82 340 58 170.31 -108.22 -36.07
val9c 50 92 332 44 332* 56.44 332* 71.19 332* 12 332* 69 389 51 175.60 -330.71 -27.56
val9d 50 92 382 44 409 353.28 391 211.13 391 6 399 50 434 44* 134.13 -106.55 -17.76
val10a 50 97 428 47 428* 5.52 428* 25.48 428* 5 428* 143 449 91 203.20 -203.37 -40.67
val10b 50 97 436 47 436* 18.43 436* 4.67 436* 5 436* 111 459 77 194.28 -175.23 -35.05
val10c 50 97 446 47 451 93.47 446* 17.30 446* 7 448 93 498 66 205.89 -169.32 -24.19
val10d 50 97 524 47 544 156.31 530 215.04 528 6 537 61 595 54 149.31 -245.01 -40.83
Average 1.90% 63.87 0.61% 38.35 0.54% 5.24 0.99% 52.68% 10.69% 19.62% 76.17 -121.70 -21.85
Worst 8.57% 353.28 4.26% 215.04 4.26% 12 5.42% 204.26% 20.69% 93.62% 205.89 6.48 1.30
Optima 15 22 22 18 3 1 10
CPU times in seconds on a 1.8 GHz Pentium-IV PC. Times for CARPET and MA have been scaled. See section 4.4.1 for details.
Table 7: Detailed results of MO9 for egl files
File n m τ LB1 LB2 CARPET MA Time BKS |F | f1lef t f2lef t f1right f2right Time µ µ̄
egl-e1-a 77 98 51 3515 820 3625 3548 74.26 3548 4 3548 943 3824 820* 31.39 -1431.21 -357.80
egl-e1-b 77 98 51 4436 820 4532 4498 69.48 4498 3 4525 839 4573 820* 25.70 -753.88 -251.29
egl-e1-c 77 98 51 5453 820 5663 5595 71.18 5595 2 5687 836 5764 820* 23.33 -177.03 -88.51
egl-e2-a 77 98 72 4994 820 5233 5018 152.58 5018 5 5018 953 6072 820* 77.93 -1269.08 -253.82
egl-e2-b 77 98 72 6249 820 6422 6340 153.41 6340 6 6411 864 6810 820* 65.47 -1581.27 -263.54
egl-e2-c 77 98 72 8114 820 8603 8415 129.63 8395 4 8440 854 8651 820* 56.02 -230.42 -57.61
egl-e3-a 77 98 87 5869 820 5907 5898 242.00 5898 12 5956 917 7935 820* 149.30 -1860.85 -155.07
egl-e3-b 77 98 87 7646 820 7921 7822 255.35 7816 7 7911 872 8455 820* 104.96 -1592.97 -227.57
egl-e3-c 77 98 87 10019 820 10805 10433 206.35 10369 4 10349 864 10511 820* 95.80 -827.25 -206.81
egl-e4-a 77 98 98 6372 820 6489 6461 291.87 6461 11 6548 890 7362 820* 166.66 -2003.50 -182.14
egl-e4-b 77 98 98 8809 820 9216 9021 312.85 9021 5 9116 874 9584 820* 145.44 -389.35 -77.87
egl-e4-c 77 98 98 11276 820 11824 11779 252.38 11779 1 11802 820* 11802 820* 111.18 -1041.39 -1041.39
egl-s1-a 140 190 75 4992 912 5149 5018 208.61 5018 11 5102 1023 6582 924 91.23 -1895.27 -172.30
egl-s1-b 140 190 75 6201 912 6641 6435 208.77 6435 7 6500 984 8117 912* 83.35 -1486.99 -212.43
21
egl-s1-c 140 190 75 8310 912 8687 8518 165.55 8518 5 8694 946 9205 912* 71.72 -1413.13 -282.63
egl-s2-a 140 190 147 9780 979 10373 9995 874.36 9995 13 10207 1058 12222 979* 497.49 -1233.84 -94.91
egl-s2-b 140 190 147 12886 979 13495 13174 760.50 13174 10 13548 1058 14334 979* 516.85 -496.91 -49.69
egl-s2-c 140 190 147 16221 979 17121 16795 746.93 16715 4 16932 1040 16975 979* 399.14 -4042.88 -1010.72
egl-s3-a 140 190 159 10025 979 10541 10296 1070.50 10296 10 10456 1099 12605 979* 699.12 -3899.11 -389.91
egl-s3-b 140 190 159 13554 979 14291 14053 1064.01 14028 6 14004 1040 15103 979* 609.54 -6439.40 -1073.23
egl-s3-c 140 190 159 16969 979 17789 17297 874.30 17297 4 17825 998 18043 979* 493.60 -748.77 -187.19
egl-s4-a 140 190 190 12027 1027 13036 12442 1537.59 12442 3 12730 1040 12912 1027* 838.24 -255.12 -85.04
egl-s4-b 140 190 190 15933 1027 16924 16531 1430.26 16531 1 16792 1027* 16792 1027* 720.31 -271.94 -271.94
egl-s4-c 140 190 190 20179 1027 21486 20832 1495.02 20832 1 21309 1027* 21309 1027* 381.86 -1425.00 -1425.00
Average 4.74% 2.47% 526.99 2.40% 5.79 3.69% 6.31% 12.94% 0.05% 268.99 -1531.94 -350.77
Worst 8.61% 4.46% 1537.59 4.46% 13 5.85% 16.22% 35.20% 1.32% 838.24 -177.03 -49.69
Nb BKS retrieved 0 19 24 2
Nb BKS improved 2
Nb Optima 3 23
CPU times in seconds on a 1.8 GHz Pentium-IV PC. Times for CARPET and MA have been scaled.
Using Lagrangean Relaxation to Minimize the Weighted
Number of Late Jobs on a Single Machine
Stéphane Dauzère-Pérès,1,* Marc Sevaux2,†
1
IRCCyN/Ecole des Mines de Nantes, La Chantrerie, BP 20722,
F-44307 Nantes Cedex 03, France
2
Université de Valenciennes, LAMIH/SP, Le Mont-Houy,
F-59313 Valenciennes Cedex, France
Received 30 April 2001; revised 21 February 2002; accepted 15 July 2002
DOI 10.1002/nav.10056
Abstract: This paper tackles the general single machine scheduling problem, where jobs have
different release and due dates and the objective is to minimize the weighted number of late jobs.
The notion of master sequence is first introduced, i.e., a sequence that contains at least an optimal
sequence of jobs on time. This master sequence is used to derive an original mixed-integer linear
programming formulation. By relaxing some constraints, a Lagrangean relaxation algorithm is
designed which gives both lower and upper bounds. The special case where jobs have
equal weights is analyzed. Computational results are presented and, although the duality gap
becomes larger with the number of jobs, it is possible to solve problems of more than 100 jobs.
© 2002 Wiley Periodicals, Inc. Naval Research Logistics 50: 273–288, 2003.
1. INTRODUCTION
A set of n jobs { J 1 , . . . , J n }, subject to release dates r i and due dates d i , have to be
scheduled on a single machine. The processing time of jobs on the machine is denoted by p i , and
a weight w i is associated with each job. The machine can only process one job at a time. A
scheduled job completed before its due date is said to be early or on time, and late otherwise.
Let U i be equal to 1 if job J i is late in a schedule, and to 0 otherwise. The objective is to
minimize the weighted number of late jobs, or equivalently to maximize the weighted number
of early jobs. A well-known and important remark is that there is always an optimal schedule
in which late jobs are sequenced after all the early jobs.
This single-machine scheduling problem, noted 1兩r j 兩¥ w j U j in the standard classification, is
strongly ᏺᏼ-hard [8]. When all weights are equal (1兩r j 兩¥ U j ), the problem remains ᏺᏼ-Hard,
but becomes polynomially solvable if all release dates are equal (1储¥ U j ) [9] (O(n log n)), or
* This paper was written while this author was invited professor at the Department of Finance and
Management Science, Norwegian School of Economics and Business Administration, Helleveien 30,
N-5035 Bergen-Sandviken, Norway.
† This research was performed while this author was at Ecole des Mines de Nantes.
Correspondence to: S. Dauzère-Pérès
© 2002 Wiley Periodicals, Inc.

274 Naval Research Logistics, Vol. 50 (2003)
if release and due dates are similarly ordered (r i ⬍ r j f d i ⱕ d j @( J i , J j )) [6] (O(n 2 )), [7]
(O(n log n)). However, some exact approaches have recently been proposed for this problem
[1, 5]. Lawler [7] showed that the Moore’s algorithm [9] could be applied when processing times
and weights are agreeable, i.e., p i ⬍ p j f w i ⱖ w j @( J i , J j ). Finally, branch-and-bound
procedures have been developed to solve the case where all release dates are equal (1储¥ w j U j )
in [12] and [11]. To our knowledge, no algorithm has been proposed to solve the general
problem 1兩r j 兩 ¥ w j U j .
In this paper, based on the notion of master sequence, i.e., a sequence from which an optimal
sequence can be extracted, a new mixed-integer linear programming formulation is introduced.
Using this formulation, a Lagrangean relaxation algorithm is derived. Lagrangean relaxation is
a powerful optimization tool from which heuristic iterative algorithms can be designed, where
both upper and lower bounds are determined at every iteration. It is thus possible to always
know the maximum gap between the best solution found and the optimal solution, and stop the
algorithm when this gap is small enough. One condition that is often associated with the
efficiency of Lagrangean relaxation approaches is to relax as few constraints as possible, in
order to obtain good bounds when solving the relaxed problem. This is why our formulation
compares very favorably to other known ones (see [4] for a study of classical formulations for
this problem). Only one constraint type, coupling variables of different jobs, needs to be relaxed
to obtain an easily solvable problem, which can be solved independently for each job. This is
not the case for other formulations, including discrete-time formulations, which often have
many coupling constraints.
The master sequence is introduced in Section 2, and the resulting mixed-integer linear
programming formulation is given and discussed in Section 3. Section 4 shows how the size of
the master sequence, and thus the size of the model, can be reduced. Section 5 presents the
Lagrangean relaxation algorithm, and Section 6 several improvements of the algorithm. The
nonweighted case is analyzed in more details in Section 7. Numerical results on a large set of
test instances are given and discussed in Section 8. Finally, some conclusions and perspectives
are drawn in Section 9.
2. THE MASTER SEQUENCE
In the remainder of this paper, because we are only interested in sequencing jobs on time (late
jobs can be set after the jobs on time), the sequence of jobs will often mean the sequence of early
jobs. Many results in this paper are based on the following theorem.
THEOREM 1: There is always an optimal sequence of jobs on time that solves the problem
1兩r j 兩 ¥ w j U j , in which every job J j is sequenced just after a job J i such that either condition (1)
d i ⬍ d j , or (2) d i ⱖ d j and r k ⱕ r j @J k sequenced before J j , holds.
PROOF: First, note that conditions (1) or (2) being satisfied is equivalent to condition (3),
d i ⱖ d j and ?J k sequenced before J j such that r k ⬎ r j , not being satisfied.
The proof goes by showing that, by construction, it is possible to change any optimal
sequence into an optimal sequence that satisfies conditions (1) or (2).
Suppose that we have a sequence in which some (or all) ready jobs do not satisfy one of the
conditions. Starting from the beginning of the sequence, find the first pair of jobs ( J i , J j ) in the
sequence that does not satisfy the two conditions, i.e., for which condition (3) holds. If t i and
t j denote the start times of the two jobs, the latter condition ensures that, after interchanging the
Dauzère-Pérès and Sevaux: Lagrangean Relaxation for Minimizing Late Jobs 275
two jobs, J j can start at t i (since ?J k sequenced before J j such that r j ⬍ r k ⱕ t i ). Hence, J i will
end at the same time as J j before the interchange (t i ⫹ p i ⫹ p j ), and thus will still be on time
(since t i ⫹ p i ⫹ p j ⱕ d j ⱕ d i ).
The interchange must be repeated if J j and the new job just before it do not satisfy conditions
(1) or (2), until one of these conditions is satisfied for J j and the job just before it, or J j is
sequenced first.
The procedure is repeated for all jobs until the conditions are satisfied for all jobs. Because
once a job has been moved, it will never go back again, one knows that the procedure will not
be repeated more than n times, i.e., takes a finite amount of time. 䊐
We will denote by ᏿ the subset of sequences in which jobs satisfy the conditions in Theorem
1. In the sequel, we will only be interested in sequences in ᏿, since we know that it always
contains an optimal sequence.
THEOREM 2: If, in a sequence of ᏿, job J j is after jobs J i such that r j ⬍ r i , then there is
at least one job J i such that d i ⬍ d j .
PROOF: By contradiction, if all jobs J k before J j such that r k ⬍ r i verify d k ⱖ d j , then

neither of the conditions (1) and (2) is satisfied. Thus, the sequence is not in ᏿. 䊐
COROLLARY 1: Consider a job J j . If, for every job J i such that r j ⬍ r i , condition d j ⱕ d i
holds, then, in every sequence of ᏿ (i.e., in an optimal sequence), job J j is sequenced before all
jobs J i .
COROLLARY 2: Consider a job J i . If, for every job J j such that d j ⬍ d i , condition r j ⱕ r i
holds, then, in every sequence of ᏿ (i.e., in an optimal sequence), job J i is sequenced after all
jobs J j .
We want to show that it is possible to derive what will be called a master sequence,
denoted by ␴, which “contains” every sequence in ᏿. A master sequence is a sequence in
which each job may appear several times. A sequence in ᏿ is “created” from the master
sequence by selecting at most one position for each job. The largest possible master sequence
is ( J 1 , J 2 , . . . , J n , J 1 , J 2 , . . . , J n , . . . , J 1 , J 2 , . . . , J n ), where ( J 1 , J 2 , . . . , J n ) is repeated
n times, and thus has n 2 positions (this number is reduced in the sequel). Corollary 1 implies that
there is only one position for J j in the master sequence, and Corollary 2 that there is only one
position for J i .
THEOREM 3: In a master sequence, if r i ⬍ r j and d i ⬎ d j , then there is a position for J i

before J j and a position for J i after J j .
PROOF: Because r i ⬍ r j , Condition (2) in Theorem 1 is satisfied for the pair of jobs ( J i , J j ),
and because d i ⬎ d j , Condition (1) is satisfied for the pair ( J j , J i ). Hence, there is a position
in the master sequence for J i before and after J j . 䊐
Hence, there must be a position in a master sequence for J i after every job J j such that r i ⬍
n共n ⫹ 1兲
r j and d i ⬎ d j . This shows that there will be at most 2 positions in the master sequence.
Table 1. Data for a 5-job problem.

Jobs J1 J2 J3 J4 J5
ri 0 5 8 12 14
pi 8 6 5 6 10
di 16 26 24 22 32
COROLLARY 3: If, for every job J j such that r i ⬍ r j , the condition d i ⱕ d j holds, then there
is only one position for job J i in the master sequence.
Corollary 3 shows that, when release and due dates are similarly ordered (as in Kise, Ibaraki,
and Mine [6]), the master sequence will be the sequence of jobs in increasing order of their
release dates (or due dates if some jobs have equal release dates). In the nonweighted case (w i ⫽
1, @J i ), the problem is then polynomially solvable using the algorithm proposed in [6] (in
O(n 2 )) or in [7] (in O(n log n)).
An interesting and important property of the master sequence is a kind of transitivity property.
If job J i is set before and after J j in the master sequence because either Condition (1) or (2) of
Theorem 1 holds, and if J j is set before and after J k in the master sequence because either
Condition (1) or (2) holds, then either Condition (1) or (2) of Theorem 1 holds and J i is set
before, and after J k in the master sequence.
EXAMPLE 1: Let us consider a 5-job problem with the data of Table 1.

Considering sequences in ᏿, by Corollary 1, one knows that J 1 is set before all jobs
(conditions r 1 ⬍ r i and d 1 ⬍ d i are satisfied for every job J i ⫽ J 1 ), and all jobs are set before
J 5 (conditions r i ⬍ r 5 and d i ⬍ d 5 are satisfied for every job J i ⫽ J 5 ). Hence, in the master
sequence ␴, job J 1 will be set first and job J 5 last.
The master sequence has the following form:
␴ ⫽ 共J 1, J 2, J 3, J 2, J 4, J 3, J 2, J 5兲.
Every sequence of jobs in ᏿ can be constructed from ␴. In this example, there are numerous
sequences or early jobs (more than 40). For instance, the subset of sequences containing 5 jobs
is:
兵共J 1, J 2, J 3, J 4, J 5兲, 共J 1, J 2, J 4, J 3, J 5兲, 共J 1, J 3, J 2, J 4, J 5兲, 共J 1, J 3, J 4, J 2, J 5兲, 共J 1, J 4, J 3, J 2, J 5兲其
One can check that each of these sequences is included in ᏿.
The algorithm to create the master sequence ␴ is sketched below. We suppose that the jobs
are preordered in nondecreasing order of their release dates, and J៮ denotes the set of jobs already
sequenced. In Step 3, jobs that are already in the master sequence (set J៮ ) and have a position
after the new job J i in ␴ (i.e., satisfying the conditions of Theorem 3) are added in nondecreasing
order of their due dates. Hence, to speed up the algorithm, jobs added to J៮ in Step 2 are ordered
in nondecreasing order of their due dates.
FOR every job J i 僆 J DO
1. ␴ 4 ␴ 艛 J i.
2. J៮ 4 J៮ 艛 J i .
3. FOR every job J j 僆 J៮ , J j ⫽ J i , such that d j ⱖ d i DO ␴ 4 ␴ 艛 J j .
The algorithm has a time complexity of O(n 2 ). The job set at position k in ␴ is denoted ␴ (k).
The positions added in Step 3 of the algorithm are said to be generated by J i . In the previous
example, the 4th position of ␴ ( J 2 ) is generated by J 3 , and the 6th and 7th positions ( J 3 and
J 2 ) are generated by J 4 . The number of positions in the master sequence is denoted by P. Recall
n共n ⫹ 1兲
that P ⱕ 2 . Actually, P will only be equal to its upper bound if the job with the smallest
release date has also the largest due date, the job with the second smallest release date has the
second largest due date, and so on (see Theorem 3). This is clearly a very special case and, in
n共n ⫹ 1兲
practical experiments, P will be much smaller than 2 .
3. A NEW MIXED-INTEGER LINEAR PROGRAMMING FORMULATION

The variable u k is equal to 1 if the job ␴ (k) is sequenced on time at position k in the master
sequence, and is equal to 0 otherwise. Let us denote by t k the start time of the job at position
k in the master sequence (t k is fictitious if u k ⫽ 0). Then, one can derive the following model:
冦
冘wU,
n
c* ⫽ min i i 共1兲
i⫽1
tk ⫺ tk⫺1 ⫺ p␴ 共k⫺1兲 uk⫺1 ⱖ 0 k ⫽ 2, . . . , P, 共2兲
tk ⫺ r␴ 共k兲 uk ⱖ 0 ᭙ k, 共3兲
tk ⫹ p␴ 共k兲 uk ⫺ d␴ 共k兲 ⫺ Dk 共1 ⫺ uk 兲 ⱕ 0 ᭙ k, 共4兲
冘 u ⫹U ⫽1
P
k i ᭙ i, 共5兲
k⫽1
␴ 共k兲⫽i
uk 僆兵0, 1其 ᭙ k, 共6兲
Ui 僆兵0, 1其 ᭙ i, 共7兲
where D k is chosen big enough to not constrain the jobs sequenced before k, for instance,
Dk ⫽ max
l⫽1, . . . ,k⫺1
d ␴ 共l兲 ⬎d ␴ 共k兲
共d␴ 共l兲 ⫺ d␴ 共k兲兲冉⫽ max
l⫽1, . . . ,k⫺1
冊
共0, d␴ 共l兲 ⫺ d␴ 共k兲兲 .
By Constraint (2) we ensure that, if the job at the kth position in the master sequence is set
on time (u k ⫽ 1), then the job at position k ⫹ 1 cannot start before the completion of the job
at position k. If u k ⫽ 0, the constraint only ensures that t k⫹1 ⱖ t k . Constraint (3) specifies that,
if the job is scheduled on time, it cannot start before its release date. By Constraint (4), if the
job at position k is set on time (u k ⫽ 1), then it has to be completed before its due date. If u k ⫽
0, the constraint is redundant. Finally, Constraint (5) ensures that at most one position is used
for each job, or the job is late (U i ⫽ 1).
In the previous model, it is possible to replace Constraint (3) by t k ⫺ r ␴ (k) ⱖ 0 [or

equivalently to remove u k from Constraint (3)]. The new constraint is numbered (3⬘). Theorem
4 will prove the validity of the resulting model.
In the nonweighted case (w j ⫽ 1, @J j ), if Constraint (4) is replaced by t k ⫹ p ␴ (k) u k ⫺
d ␴ (k) ⱕ 0 [or equivalently D k ⫽ 0 in Constraint (4)], then the resulting formulation still
provides an optimal solution to the problem. The new constraint is numbered (4⬘). Although the
nonweighted case will be analyzed in more details in Section 7, the following theorem is
introduced here because it is also useful for the weighted case.
THEOREM 4: In the nonweighted case, there is always an optimal sequence of ᏿ that

satisfies Constraints (2), (3⬘), (4⬘), and (5)–(7).
PROOF: The proof goes by showing that the only case where there is a problem is when J j
can be sequenced before and after J i in the master sequence, and r j ⬍ r i and d j ⬎ d i , and J i
is not sequenced in the optimal sequence. It can be shown that Constraints (2), (3⬘), and (4⬘)
prevent job J j to start between d i ⫺ p j [Constraint (4⬘)] and r i [Constraint (3⬘)]. This is only a
problem if d i ⫺ p j ⬍ r i . If this is the case, then p i ⬍ p j (since J i is not late if started at its release
date r i ). Hence, in an optimal solution where J j starts in the interval [d i ⫺ p j , r i ], i.e., ends in
the interval [d i , r i ⫹ p j ], J j can be replaced by J i , and the sequence will remain optimal since
J i starts after r i and ends before d i . 䊐
The proof of Theorem 4 is based on equal weight for jobs. In the weighted case, following
the proof of Theorem 4, D k can be chosen as follows:
Dk ⫽ max 共0, r␴ 共l兲 ⫺ d␴ 共k兲兲

l⫽1, . . . ,k⫺1
Hence, the case where d i ⫺ p j ⬍ r i , discussed in the proof of Theorem 4, is avoided. In

numerical experiments, D k is very often equal to zero.
4. REDUCING THE MASTER SEQUENCE

Because the size of the model is directly linked to the length of the master sequence, it is
desirable to remove as many positions as possible from ␴. Not only solution procedures will be
more efficient, but the model will be tighter and will give better lower bounds by Lagrangean
relaxation.
By Constraints (2) and (3), t k ⱖ maxl⫽1, . . . ,k⫺1 r ␴ (l ) . Hence, the first reduction will be done
by removing positions k such that maxl⫽1, . . . ,k⫺1 r ␴ (l ) ⫹ p ␴ (k) ⬎ d ␴ (k) .
Several dominance rules are proposed in [5] for the nonweighted case. However, if parameter
D k is changed according to Theorem 4, all of them do not apply. This is because, in the resulting
formulation, when job J j is before and after J i in the master sequence and J i is late, the position
of J j after J i might need to be occupied in an optimal solution. One could show that this is not
the case with the initial formulation. Our preliminary numerical experiments showed that
reducing parameter D k was more important than using the lost dominance rules.
We will describe here the dominance rules that still apply to our formulation, and which have
been modified for the weighted case (see [5] for details).
In the master sequence, if Conditions (1) r i ⬍ r j , (2) r i ⫹ p i ⱖ r j ⫹ p j , (3) r i ⫹ p i ⫹ p j ⬎
d j , (4) r j ⫹ p j ⫹ p i ⬎ d i , (5) d i ⫺ p i ⱕ d j ⫺ p j , and (6) w j ⱕ w i hold, then J j dominates
J i and all positions of job J i can be removed from the master sequence. Because of Conditions
(3) and (4), only one of the two jobs can be scheduled on time. In an optimal solution, either
both jobs are late, or it is always possible to find a solution in which job J j is on time and the
total weight of late jobs is as small as a solution with job J i on time.
Another dominance rule is based on the fact that if there is a position l and a job J j [ J j ⫽
␴ (l )] such that Conditions (1) r ␴ (l ) ⫹ p ␴ (l ) ⱖ r j ⫹ p j , (2) p ␴ (l ) ⱖ p j , (3) r ␴ (l ) ⫹ p ␴ (l ) ⫹ p j ⬎
d j , (4) r j ⫹ p j ⫹ p ␴ (l ) ⬎ d ␴ (l ) , (5) d ␴ (l ) ⫺ p ␴ (l ) ⱕ d j ⫺ p j , and (6) w ␴ (l ) ⱖ w j are satisfied,
then J j dominates position l, and thus the latter can be removed. This is because, if there is an
optimal solution in which position l is occupied (i.e., job J ␴ (l ) is on time), then, by Condition
(3), J j is late. The solution can be changed to another optimal solution in which J ␴ (l ) is replaced
by J j .
5. A LAGRANGEAN RELAXATION ALGORITHM
Following Theorem 4 and the remarks in Section 3, the mixed-integer linear programming
formulation is now
冦
冘wU,
n
c* ⫽ min i i 共8兲
i⫽1
tk ⫺ r␴ 共k兲 ⱖ 0 ᭙ k, 共10兲
冘 u ⫹U ⫽1
P
k i ᭙ i, 共12兲
k⫽1
␴ 共k兲⫽i
Ui 僆兵0, 1其 ᭙ i. 共14兲
By relaxing Constraint (9) using Lagrangean multipliers ␭ k (k ⫽ 2, . . . , P), the model

becomes
冋冘册
冦
冘 ␭ 共t ⫺ t
n P
max min wi Ui ⫺ k k k⫺1 ⫺ p␴ 共k⫺1兲 uk⫺1 兲 , 共15兲

␭ k ⱖ0 t k ,u k ,U i i⫽1 k⫽2
冘 u ⫹U ⫽1
P
k i ᭙ i, 共12兲
k⫽1
␴ 共k兲⫽i
To use Lagrangean relaxation, one needs to solve the above model for given values of ␭ k (k ⫽
2, . . . , P). The objective function can be written:
min
t k ,u k ,U i
冋冘 n
i⫽1
wi Ui ⫹ 冘␭ p
P
k⫽2
k u
␴ 共k⫺1兲 k⫺1 ⫹ ␭2 t1 ⫹ 冘共␭
P⫺1
k⫽2
k⫹1 ⫺ ␭k 兲tk ⫺ ␭P tP 册 (16)
Because Constraint (9) has been relaxed, variables t k are now independent and bounded
through Constraints (10) and (11). Hence, if the coefficient of t k ( ␭ k⫹1 ⫺ ␭ k ) is positive, t k will
be chosen as small as possible to minimize the cost, i.e., r ␴ (k) [because of (10)], and if the
coefficient is negative, t k will be chosen as large as possible, i.e., d ␴ (k) ⫹ D k ⫺ ( p ␴ (k) ⫹ D k )u k
P
[because of (11)]. Moreover, using (12), U i can be replaced by 1 ⫺ ¥ k⫽1 u k in the criterion.
␴ 共k兲⫽i
Hence, (16) becomes
min
uk
冋冉
冘w 1⫺ 冘 u ⫹ 冘␭ p
n
i⫽1
i
k⫽1
␴ 共k兲⫽i
P
k 冊 P
k⫽2
k ␴ 共k⫺1兲 uk⫺1 ⫹ ␭2 r␴ 共1兲 ⫹ 冘
P⫺1
k⫽2
共 ␭ k⫹1 ⫺ ␭ k 兲ⱖ0
共␭k⫹1 ⫺ ␭k 兲r␴ 共k兲
⫹ 冘
P⫺1
k⫽2
共 ␭ k⫹1 ⫺ ␭ k 兲⬍0
共␭k⫹1 ⫺ ␭k 兲共d␴ 共k兲 ⫹ Dk ⫺ 共p␴ 共k兲 ⫹ Dk 兲uk 兲 ⫺ ␭P 共d␴ 共P兲 ⫹ DP ⫺ 共p␴ 共P兲 ⫹ DP 兲uP 兲册
Note that the minimization now only depends on variables u k . Since r i and d i are data, several
terms of the previous expression can be ignored in the optimization:
min
uk
冘冘
n
i⫽1
冋 P
k⫽1; ␴ 共k兲⫽i
共 ␭ k⫹1 ⫺ ␭ k 兲ⱖ0
共␭k⫹1 pi ⫺ wi 兲uk ⫹ 冘
P
共 ␭ k⫹1 ⫺ ␭ k 兲⬍0
共␭k⫹1 pi ⫺ 共␭k⫹1 ⫺ ␭k 兲共pi ⫹ Dk ⫺ wi 兲兲uk 册
or, after simplification,
min
uk
冘冘
i⫽1
n
冋 P
共 ␭ k⫹1 ⫺ ␭ k 兲ⱖ0
共␭k⫹1 pi ⫺ wi 兲uk ⫹ 冘
P
共 ␭ k⫹1 ⫺ ␭ k 兲⬍0
共␭k pi ⫺ 共␭k⫹1 ⫺ ␭k 兲Dk ⫺ wi 兲uk , 册 (17)
where ␭1 and ␭ P⫹1 are parameters such that ␭ 1 ⫽ ␭ P⫹1 ⫽ 0.

To minimize the cost, and to satisfy Constraint (12), one has to determine, for every job J i ,
the position k⬘ such that ␴ (k⬘) ⫽ i with the smallest coefficient in (17), i.e., ( ␭ k⫹1 p i ⫺ w i ) or
( ␭ k p i ⫹ ( ␭ k⫹1 ⫺ ␭ k ) D k ⫺ w i ), depending on the sign of ( ␭ k⫹1 ⫺ ␭ k ). If the coefficient is
positive, then u k ⫽ 0 @k such that ␴ (k) ⫽ i, and U i ⫽ 1, and if the coefficient is negative,
then u k⬘ ⫽ 1, u k ⫽ 0 @k ⫽ k⬘ such that ␴ (k) ⫽ i, and U i ⫽ 0.
REMARK 1: We shall show that solving the relaxed problem can be done in ᏻ(P) time.
The solution of the relaxed problem would be the same, i.e., integral, if Constraints (13) and
(14) were to be replaced by 0 ⱕ u k ⱕ 1, @k. Hence, the Lagrangean relaxation bound is
identical to the bound obtained by linear relaxation (see Parker and Rardin [10]). However, this
bound can be determined much faster, because every subproblem can be trivially solved, and
larger problems can be solved. Actually, before implementing the Lagrangean relaxation
algorithm, we performed some preliminary testing using linear relaxation with a standard and
efficient LP package. The quality of the bound was better than all other formulations we had
tested before, including time-indexed formulations (see [4]).
It is relatively easy to interpret the impact of the values of ␭ k , p i , or w i . Increasing ␭ k will
force the associated Constraint (9) to be satisfied, i.e., t k to be chosen as large as possible and
equal to d ␴ (k) ⫹ D k ⫺ ( p ␴ (k) ⫹ D k )u k (u k to 0), and t k⫺1 as small as possible and equal to
r ␴ (k⫺1) . Intuitively, a job with a large processing time that is set on time might force more jobs
to be late than a job with a smaller processing time. Hence, it is natural to favor jobs with small
processing times. This is consistent with (17), where the coefficient of u k will increase with
p ␴ (k) , and then has more chances to become positive, thus inducing u k ⫽ 0, i.e., job J ␴ (k) is not
set in position k. The exact opposite can be said about the weight, since the larger its weight,
the more you want to sequence a job. Again, this is in accordance with (17), where the
coefficient of u k will decrease with w ␴ (k) , and then has more chances to become negative, thus
inducing u k ⫽ 1, i.e., job J ␴ (k) is set in position k.
The following algorithm is proposed to solve our problem using Lagrangean relaxation and
subgradient optimization (see Parker and Rardin [10]).
p␴ 共k兲
Step 1—Initialization of the Lagrangean variables ␭ k : ␭ 0k ⫽ f n ⴱ pmax ⴱ wmax ⴱ w␴共k兲 @k (where p max ,
resp. w max , is the largest processing time, resp. weight, among all jobs, and f a parameter),
and s ⫽ 0 (where s denotes the iteration counter).
Step 2—Initialize the various parameters: U i ⫽ 1, coef(i) ⫽ ⬁ and pos(i) ⫽ ⫺1 @i, u k ⫽
0 @k, and ␭ s1 ⫽ ␭ P⫹1s
⫽ 0.
Step 3—Solve the relaxed problem:
Step 3.1—For k ⫽ 1, . . . , P, if ␭ sk⫹1 ⫺ ␭ sk ⱖ 0 then tempcoef ⫽ ␭ k⫹1 s
p i ⫺ w i , else
tempcoef ⫽ ␭ k p i ⫺ ( ␭ k⫹1 ⫺ ␭ k ) D k ⫺ w i .
s s s
If tempcoef ⬍ coef( ␴ (k)), then coef( ␴ (k)) ⫽ tempcoef and pos( ␴ (k)) ⫽ k.
Step 3.2—For i ⫽ 1, . . . , n, if coef(i) ⱕ 0 then u pos(i) ⫽ 1 and U i ⫽ 0.
Step 4—Compute the lower bound:
LB ⫽ 冘
n
i⫽1
冋 wi ⫹ 冘
P
k⫽1;␴共k兲⫽i
I1共 ␭ ks, ␭ k⫹1
s
册
, u k兲 ,
where I1( ␭ sk , ␭ k⫹1

s
, u k ) ⫽ ( ␭ k⫹1
s
⫺ ␭ sk )r i ⫹ ( ␭ sk⫹1 p i ⫺ w i )u k if ␭ sk⫹1 ⫺ ␭ sk ⱖ 0, and I1( ␭ sk ,
␭ k⫹1 , u k ) ⫽ ( ␭ k⫹1 ⫺ ␭ k )(d i ⫹ D k ⫺ D k u k ) ⫹ ( ␭ sk p i ⫺ w i )u k otherwise.
s s s
Step 5—Compute an upper bound by sequencing as many jobs as possible among the jobs J i
that are set on time in the solution associated to the lower bound, i.e., such that U i ⫽ 0.
Step 6—Update the Lagrangean variables ␭ k :
␭ s⫹1
k 冉
⫽ max 0, ␭sk ⫺ ␳s
tk ⫺ tk⫺1 ⫺ p␴ 共k⫺1兲 uk⫺1
兩tk ⫺ tk⫺1 ⫺ p␴ 共k⫺1兲 uk⫺1 兩
, 冊
where t k ⫽ r ␴ (k) if ( ␭ k⫹1
s
⫺ ␭ sk ) ⱖ 0, and t k ⫽ d ␴ (k) ⫹ D k ⫺ (D k ⫹ p ␴ (k) )u k otherwise.
Update ␳ s⫹1 , and s ⫽ s ⫹ 1.
Step 7—If no stopping conditions are met, go to Step 2.
We use a simple and fast greedy algorithm to determine the upper bound in Step 5. From k ⫽
1 to k ⫽ P, job J ␴ (k) is added to the sequence of early jobs if u k ⫽ 1 and J ␴ (k) is on time. The
finishing time of the current sequence is updated each time a new job is added.
Various parameters have to be initialized and adjusted to ensure the best convergence of the
algorithm for different types of instances. After sd iterations without improvement, the param-
eter ␳ s is decreased by a factor of 100 ⫻ (1 ⫺ red ␳ )%. Various stopping conditions are
checked: Maximum number of iterations IterMax, step ␳ smaller than or equal to ␳ min , and of
course if the optimum is found, i.e., the lower and upper bounds are equal. The parameters could
be adjusted to improve the results on some instances, but we have decided to use generic
parameters instead. After some preliminary testing, we chose the following values: f ⫽ 0.4,
␳ 0 ⫽ 1.6, sd ⫽ 40, and red ␳ ⫽ 0.9. For the stopping conditions, we used IterMax ⫽
100,000 and ␳ min ⫽ 10 ⫺5 . Actually, in our numerical experiments, the number of iterations is
never larger than 20,000.
As already shown, every relaxed problem in Step 3 is solved very quickly, in ᏻ(P) time where
n共n ⫹ 1兲
P is not larger than 2 . Hence, many iterations can be performed, even for large instances.
6. IMPROVING THE ALGORITHM

Several improvements are proposed. The first one is based on a rewriting of the formulation.
In the model, because of Constraint (9), Constraint (10) can be rewritten as
t k ⫺ rr k ⱖ 0 ᭙ k,
where rr k ⫽ maxl⫽1, . . . ,k r ␴ (l ) are release dates per position. To include this change in the
algorithm, it suffices to replace r ␴ (k) by rr k .
A similar rewriting can be performed for Constraint (11) when D k ⫽ 0 @k, as follows:
t k ⫹ p ␴共k兲u k ⫺ dd k ⱕ 0,
where dd k ⫽ minl⫽k, . . . ,P;D k⫽0 d ␴ (l ) are due dates per position. This rewriting can be done for
all positions in the nonweighted case since D k ⫽ 0 @k.
Although they do not improve the lower bound obtained by linear relaxation, and thus by
Lagrangean relaxation, these changes often considerably speed up the algorithm by better
updating the Lagrangean multipliers in Step 6. This is because the positions for a job are better
differentiated whereas, in the original formulation, they all have similar Constraints (10). Hence,
the algorithm will more quickly choose the best position(s) for a job, and will require less
iterations to converge to the lower bound.
Another improvement uses the following property to tighten Constraint (9) in the model.
THEOREM 5: Let J i and J j be two jobs in a master sequence. If a position k of J i is generated

by J j , then there is an optimal schedule in which k is either not occupied or occupied and such
that t k ⱖ d j ⫺ p i .
PROOF: We want to prove that if, in an optimal schedule S, the position k of J i generated
by J j is occupied and t k ⱕ d j ⫺ p i , then this schedule can be transformed into an equivalent
optimal schedule S⬘ in which J i is sequenced before J j (i.e., position k is not occupied).
Since a position of J i is generated by J j , we know that r i ⬍ r j and d i ⬎ d j . Hence, moving

J i before J j will just shift J j and the jobs between J j and J i in S by p i and, because t k ⱕ d j ⫺
p i in S, the completion time of the shifted jobs will not be larger than d j . By definition of the
master sequence, and because position k is generated by J j (see the end of Section 2 for the
definition of a generated position), the due dates of the jobs between J j and J i in S are larger than
or equal to d j . Thus, the schedule S⬘ is feasible. 䊐
Following Theorem 5, Constraints (10) can be tightened (the added term is positive) for
generated positions k (i.e., such that rr k ⬎ r ␴ (k) ) as follows:
t k ⫺ rr k ⫺ RR ku k ⱖ 0 ᭙k
where RR k ⫽ max(0, d ␴ (l ) ⫺ p ␴ (k) ⫺ rr k ), and l is the position that generated position k (i.e.,
r ␴ (k) ⬍ r ␴ (l ) and d ␴ (k) ⬎ d ␴ (l ) ).
The relaxed problem in the Lagrangian relaxation changes accordingly by adding the new
term in the objective function, and by considering the coefficient ( ␭ k⫹1 p i ⫹ ( ␭ k⫹1 ⫺
␭ k ) RR k ⫺ w i ) when ( ␭ k⫹1 ⫺ ␭ k ) is positive. Strengthening the constraints helps to improve
the quality of the lower bound. Moreover, it also accelerates the algorithm by again better
differentiating the positions.
7. THE NONWEIGHTED CASE
The mixed-integer linear programming model defined in Section 3 can be enhanced for the
nonweighted case, i.e., w i ⫽ 1 @i following Theorem 4 in Section 3. The new model is given
below:
冦
冘U,
n
c* ⫽ min i 共18兲
i⫽1
tk ⫹ p␴ 共k兲 uk ⫺ d␴ 共k兲 ⱕ 0 ᭙ k, 共21兲
冘 u ⫹U ⫽1
P
k i ᭙ i, 共22兲
k⫽1
␴ 共k兲⫽i
Because w i ⫽ 1 @J i and D k ⫽ 0, @k, the objective function (17) can equivalently be written
as
冘冘共max共␭ , ␭
n P
min k k⫹1 兲pi ⫺ 1兲uk . (25)

uk k⫽1;
i⫽1 ␴ 共k兲⫽i
REMARK 2: In the nonweighted case, for a given job J i , finding the position k⬘, ␴ (k⬘) ⫽
i, with the smallest coefficient in (17) is equivalent to finding the position with the smallest
coefficient ␭ k⫹1 or ␭ k , depending on the sign of ( ␭ k⫹1 ⫺ ␭ k ).
In the Lagrangean relaxation algorithm described in Section 5, the following steps are
modified:
Step 3.1—For k ⫽ 1, . . . , P, if ␭ k⫹1

s
⫺ ␭ sk ⱖ 0, then tempcoef ⫽ ␭ sk⫹1 p i ⫺ 1; else
tempcoef ⫽ ␭ k p i ⫺ 1.
s
If tempcoef ⬍ coef( ␴ (k)), then coef( ␴ (k)) ⫽ tempcoef and pos( ␴ (k)) ⫽ k.
Step 4—Compute the lower bound:
冘冘
n P
LB ⫽ n ⫹ I2共 ␭ ks, ␭ k⫹1

s
, u k兲
i⫽1 k⫽1;␴共k兲⫽i
where I2( ␭ sk , ␭ k⫹1

s
, u k ) ⫽ ( ␭ sk⫹1 ⫺ ␭ sk )r i ⫹ ( ␭ k⫹1
s
p i ⫺ 1)u k if ␭ sk⫹1 ⫺ ␭ sk ⱖ 0, and I2( ␭ sk ,
␭ k⫹1 , u k ) ⫽ ( ␭ k⫹1 ⫺ ␭ k )(d i ⫺ p i u k ) ⫹ ( ␭ k p i ⫺ 1)u k otherwise.
s s s s
Moreover, Kise, Ibaraki, and Mine’s algorithm [6] can be used to compute the upper bound
associated with the current value of the multipliers ␭ s in Step 5. This is because, when the
sequence in which jobs can be sequenced is fixed, i.e., for a given permutation of the jobs, the
optimal sequence of early jobs can be found using Kise et al.’s algorithm. In our case, the set
of jobs from which jobs have to be sequenced is the set of jobs J i such that U i ⫽ 1, and the
fixed sequence is given by the positions k such that u k ⫽ 1.
It is better to adjust the parameters for the algorithm when w i ⫽ 1, @i. After multiple trials,
we decided to use the following for all tested instances: f ⫽ 0.4, ␳ 0 ⫽ 0.6, sd ⫽ 40, and
red ␳ ⫽ 0.9. The same parameters are kept for the stopping conditions (IterMax ⫽ 100,000
and ␳ min ⫽ 10 ⫺5 ).
8. COMPUTATIONAL RESULTS
Many test problems were generated to evaluate our algorithm. For each value of n, the
number of jobs, 160 instances were randomly generated. The test program, written in C, was run
on a SUN UltraSparc workstation.
The Lagrangean relaxation algorithm was first compared with solutions obtained by using the
new formulation in the state of the art mixed-integer linear programming (MIP) solver ILOG-
CPLEX. However, even for “small” size problems of 80 jobs, very large CPU times were
observed with the MIP solver in many cases. For example, on a given instance, the search was
stopped after 19 hours with the optimal solution, but its optimality was not proved. If the
accuracy level is reduced to the one of the Lagrangean relaxation algorithm (see results in the
sequel), the CPU time of the MIP solver decreases but remains very large.
Random generator. For each job J i , a processing time p i is randomly generated from the
uniform distribution [1, 100] and a weight w i is generated from the uniform distribution [1, 10].
As in [3], two parameters K 1 and K 2 are used, and taken in the set {1, 5, 10, 20}. Because we
want data to depend on the number of jobs n, the release date r i is randomly generated from the
Table 2. Results of the nonweighted case.

Optimum CPU time (s) Gap
No. of Gap (%)
jobs No. (%) Mean St dev Max Mean St dev Max mean
n ⫽ 20 88 55.0% 2.67 1.55 8.06 0.53 0.66 3 2.65
n ⫽ 40 60 37.5% 11.26 8.80 42.73 0.87 0.87 4 2.18
n ⫽ 60 25 15.6% 29.55 20.60 106.43 1.93 1.64 9 3.21
n ⫽ 80 11 6.9% 54.32 35.95 143.96 2.76 2.05 10 3.45
n ⫽ 100 2 1.3% 85.66 57.45 230.73 3.46 2.45 11 3.46
n ⫽ 120 2 1.3% 127.01 84.87 350.26 4.02 2.92 16 3.35
n ⫽ 140 1 0.6% 184.07 121.45 497.96 4.56 3.52 15 3.26
uniform distribution [0, K 1 n], and the due date from the uniform distribution [r i ⫹ p i , r i ⫹
p i ⫹ K 2 n]. The algorithm was tested for n 僆 {20, 40, 60, 80, 100, 120, 140}. For each
combination of n, K 1 , and K 2, 10 instances are generated, i.e., 160 instances for each value of n.
Results of the nonweighted case. The Lagrangean relaxation algorithm was first run on the 1兩r j 兩
¥ U j problem. In Table 2, results are reported for each value of n. The optimum is considered
to be found when lower and upper bounds are equal. For n ⫽ 60, 25 out of 160 instances are
optimally solved, i.e., 15.6%. The CPU time necessary to find the best bounds is also reported.
For n ⫽ 80, the mean CPU time is less than 1 minute. To evaluate the efficiency of both bounds,
the gap between the upper and lower bounds is also measured and reported in the last four
columns of the table. This gap is expressed in number of jobs and in percentage in the last
column. For n ⫽ 100, the average gap is 3.46 jobs. The standard deviation and maximum gap
are also given in the table.
The results are good, and the average duality gap remains stable when n increases. Remember
that we decided to use the same parameters for our algorithm for every test instance, indepen-
dently of n, K 1 , or K 2 . The algorithm does not perform so well when the master sequence is
long. Looking at Theorem 3, this happens when there are many pairs of jobs ( J i , J j ) such that
r i ⬍ r j and d i ⬎ d j . This is the case when K 2 is large, and even more when K 1 is also small.
The same analysis holds for the CPU time, since the time to solve the relaxed problem at every
Table 3. Sensitivity of the results to parameter K 2 .

Length of ␴ CPU time (s) Gap
No. of Value Gap (%)
jobs of K 2 Mean St dev Mean St dev Mean St dev mean
n ⫽ 100 1 97.38 21.40 18.77 7.20 2.70 1.68 2.70
5 667.65 225.48 63.13 21.99 1.57 0.84 1.57
10 1161.12 313.23 104.58 30.35 3.00 1.20 3.00
20 1677.60 303.77 156.18 37.74 6.55 2.29 6.55
n ⫽ 120 1 148.05 34.45 28.21 14.55 2.70 2.04 2.25
5 1002.65 341.77 94.84 32.98 2.17 0.93 1.81
10 1716.35 462.34 154.90 40.73 3.80 1.32 3.17
20 2459.25 436.33 230.08 60.36 7.40 3.31 6.17
n ⫽ 140 1 214.90 65.15 41.90 21.76 2.67 1.93 1.91
5 1439.20 517.32 139.62 52.36 2.42 1.06 1.73
10 2412.40 677.09 226.25 66.99 4.25 2.12 3.04
20 3357.38 606.50 328.48 80.37 8.90 3.65 6.36
Table 4. Comparing with the optimal solution.

Lagrangean lower bound Lagrangean upper bound
No. of Opt. Gap Opt. (%) Opt. Gap Opt. (%)
jobs found w/mean max found w/mean max
n ⫽ 20 54.37% 2.56 15.00 98.12% 0.09 5.00
n ⫽ 40 42.50% 1.95 7.50 91.25% 0.22 2.50
n ⫽ 60 24.38% 2.44 10.00 62.50% 0.77 6.67
n ⫽ 80 15.00% 2.28 7.50 43.75% 1.17 6.25
n ⫽ 100 8.75% 2.11 8.00 34.38% 1.34 6.00
n ⫽ 120 6.25% 1.97 7.50 26.25% 1.38 9.17
n ⫽ 140 3.75% 1.86 6.43 28.12% 1.40 7.14
iteration directly depends on the length of the master sequence P. This is why the CPU time
average and standard deviation increase with the number of jobs and the value of K 2 . Table 3
reports the results and the length of the master sequence for n 僆 {100, 120, 140} and K 2 僆
{1, 5, 10, 20}. Note that, for K 2 ⫽ 20, the mean CPU time is more than three times larger
and the mean gap is approximately two times larger than in Table 2.
In [5], we propose a branch-and-bound procedure which is only valid for the nonweighted
problem. This exact method also uses the notion of master sequence, and has been tested on the
same set of instances. In a maximum running time of 1 hour, more than 95% of 140-job
instances are solved to optimality. Hence, it is possible to compare the bounds given by our
Lagrangean relaxation algorithm to the optimal solution for test instances that are optimally
solved by our exact procedure. In Table 4, we compare the two bounds for instances of more
than 80 jobs with the optimal solution.
For both the lower and upper bounds, the results are reported as follows: The first column
gives the percentage of cases where the bound and the optimal solution are equal, and the next
two columns give the mean and the maximum of the gap between the bound and the optimal
solution. Even for the largest instances (n ⫽ 140), the upper bound is very good on average,
with an error of less than 1.5%.
Better results could be obtained for the number of upper bounds equal to the optimal solution
by adjusting the parameters of the Lagrangean algorithm. We did it using the following
parameters f ⫽ 0.4, ␳ 0 ⫽ 0.05, sd ⫽ 60 and red ␳ ⫽ 0.92. For example, with n ⫽ 80, the
upper bound is equal to the optimal solution for 84.1% of instances, for n ⫽ 100, 68.2%, for
n ⫽ 120, 35.5%, and for n ⫽ 140, 26.9%. Even if these results are better, the mean gap
increases to more than 2.7% for n ⫽ 140 and the maximum gap increases to more than 14%!
Using these new parameters, the number of optimal solutions also increases. Table 5 gives the
percentage of optimum values found for different values of n. But to obtain these results, the
average CPU time also increases considerably (almost 6 minutes for n ⫽ 140).
The mean CPU times of the Lagrangean relaxation algorithm and the exact procedure are not
very different, but the standard deviation is much larger in the second approach. Recall also that
the branch-and-bound procedure can only solve nonweighted problems.
Results of the weighted case. Weights are randomly generated in the interval [1, 10]. Results
are reported in Table 6. The Lagrangean relaxation algorithm seems to be as efficient as in the
Table 5. Optimum values found.
No. of jobs n 20 40 60 80 100 120 140
Opt. found (%) 53.1 46.9 41.3 34.4 16.3 6.9 5.0
Table 6. Results of the weighted case.

CPU time (s) Gap
No. of Gap (%)
jobs Mean St dev Max Mean St dev Max mean
n ⫽ 20 4.18 2.28 12.46 4.46 3.54 20 4.07
n ⫽ 40 13.20 8.58 37.88 7.65 7.65 31 3.36
n ⫽ 60 30.66 20.39 83.28 10.26 6.37 41 3.09
n ⫽ 80 56.49 38.79 184.43 11.40 7.31 36 2.56
n ⫽ 100 86.21 58.83 231.57 14.82 8.78 47 2.70
n ⫽ 120 130.96 90.47 378.59 18.17 11.49 72 2.74
n ⫽ 140 184.66 130.34 496.85 29.18 21.46 127 3.82
nonweighted case. CPU times are close because the parameters sd and red ␳ are the same and
control the convergence of the algorithm.
Again, the last column of Table 6 gives the gap between the two bounds expressed in %. One
can notice that the gap remains very stable when the number of jobs increases. This gap can also
be compared to the one given in Table 2 and is very similar. The table equivalent to Table 3 is
not given in the weighted case, since it is very similar.
As in the nonweighted case, better results could be obtained by adjusting the parameters of
the Lagrangean algorithm. We did it again for n ⫽ 140, K 1 ⫽ 1, and K 2 ⫽ 20. The average
duality gap for the 10 instances reduces from 8.7% (max ⫽ 16.65%), when using the generic
parameters ( f ⫽ 0.4, ␳ 0 ⫽ 1.6, sd ⫽ 40 and red ␳ ⫽ 0.9), to only 2.07% (max ⫽ 4.19%)
by modifying ␳0 and sd ( ␳ 0 ⫽ 2.6 and sd ⫽ 80). To obtain these results, the average CPU time
increases to 751.09 seconds vs. 383.32 seconds with the previous parameters.
Other instances. Different ranges for weights and processing times have been considered to
generate other instances. The generic parameters used in the previous set of instances have to
be adjusted. Since the weights have a direct influence on the value of the objective function, the
initial step ␳0 in the subgradient search should be chosen accordingly. After many tests, we
decided to use ␳ 0 ⫽ w max /8.
Instances of 100 jobs are generated and solved using the previous parameters. Table 7
presents the resulting gaps, and each line corresponds to 160 test problems. The results are very
similar to the ones obtained on the previous instances. All average gaps remain below 3%.
9. CONCLUSION
This paper considers a single-machine scheduling problem in which the objective is to
minimize the weighted number of late jobs. Based on the definition of the master sequence, a
Table 7. Results with different ranges.

Gap
Range Range Gap (%)
of p i of w i Mean St dev Max mean
[1, 100] [1,10] 13.76 9.12 48 2.51
[1,99] 133.12 79.39 424 2.66
[25, 75] [1,10] 14.53 10.23 61 2.63
[1,99] 144.86 92.07 486 2.92
[1, 500] [1,10] 12.68 5.49 30 2.30
[1,99] 135.06 60.52 322 2.73
new and efficient mixed-integer linear programming formulation is derived. By relaxing some
coupling constraints using Lagrangean multipliers, the resulting problem becomes easily solv-
able. A Lagrangean relaxation algorithm is proposed and improved. Numerical experiments
have been performed on an extended set of test instances for both the nonweighted and the
weighted cases, and the algorithm performs well for problems of more than 100 jobs.
To our knowledge, our Lagrangean relaxation algorithm is the first method proposed to solve
the problem 1兩r j 兩 ¥ w j U j . We would like to improve the algorithm, in particular the number of
iterations required to obtain the lower bound, by for instance using dual ascent instead of
subgradient optimization when updating the Lagrangean multipliers.
The master sequence has also been used in a branch-and-bound method to solve the 1兩r j 兩 ¥
U j problem, i.e., the non-weighted case [5]. We plan to use the bound provided by the
Lagrangian relaxation algorithm in an exact method for the weighted case. Finally, it would be
interesting to investigate other problems where the notion of master sequence could be applied.
For instance, we believe it can be used to tackle the case where jobs can be processed in batches
(although not with families; see Crauwels, Potts, and Van Wassenhove [2]).
ACKNOWLEDGMENTS
The authors are grateful to the referees for their useful remarks and suggestions.
REFERENCES
[1] P. Baptiste, C. Le Pape, and L. Péridy, Global constraints for partial CSPs: A case study of resource
and due-date constraints, 4th Int Conf Principles Practice Constraint Program, Pisa, Italy, 1998
[Lecture Notes in Computer Science, Springer-Verlag, New York, to appear].
[2] H.A.J. Crauwels, C.N. Potts, and L.N. Van Wassenhove, Local search heuristics for single machine
scheduling with batching to minimize the number of late jobs, Eur J Oper Res 90 (1996), 200 –213.
[3] S. Dauzère-Pérès, Minimizing late jobs in the general one machine scheduling problem, Eur J Oper
Res 81 (1995), 134 –142.
[4] S. Dauzère-Pérès and M. Sevaux, Various mathematical programming formulations for a general one
machine sequencing problem, Research Report 98/3/AUTO, École des Mines de Nantes, Nantes,
France, 1998.
[5] S. Dauzère-Pérès and M. Sevaux, An exact method to minimize the number of tardy jobs in single
machine scheduling, Research Report 99/6/AUTO, École des Mines de Nantes, Nantes, France, 1999.
[6] H. Kise, T. Ibaraki, and H. Mine, A solvable case of the one-machine scheduling problem with ready
and due times, Oper Res 26(1) (1978), 121–126.
[7] E.L. Lawler, Knapsack-like scheduling problems, the Moore-Hodgson Algorithm and the “Tower of
Sets” property, Math Comput Model 20(2) (1994), 91–106.
[8] J.K. Lenstra, A.H.G. Rinnooy Kan, and P. Brucker, Complexity of machine scheduling problems,
Ann Discrete Math 1 (1977), 343–362.
[9] J.M. Moore, A n job, one machine sequencing algorithm for minimizing the number of late jobs,
Manage Sci 15(1) (1968), 102–109.
[10] R.G. Parker and R.L. Rardin, Discrete optimization, Academic Press, New York, 1988.
[11] C.N. Potts and L.N. Van Wassenhove, Algorithms for scheduling a single machine to minimize the
weighted number of late jobs, Manage Sci 34(7) (1988), 834 – 858.
[12] F.J. Villarreal and R.L. Bulfin, Scheduling a single machine to minimize the weighted number of
tardy jobs, IIE Trans 15 (1983), 337–343.
M ÉTAHEURISTIQUES :
STRAT ÉGIES POUR L’ OPTIMISATION DE LA PRODUCTION DE BIENS ET DE SERVICES
Résoudre des problèmes d’optimisation est un point clé dans l’amélioration constante de
la productivité des entreprises. Quand les méthodes traditionnelles échouent, il devient
alors naturel de se tourner vers des techniques de résolution approchée.
Les métaheuristiques jouent, aujourd’hui, un rôle primordial dans la résolution des
problèmes d’optimisation. Ces techniques sont devenues, en quelques années, des outils
incoutournables et performants. Dans cette synthèse, nous présentons un panorama des
métaheuristiques classiques (méthodes de descente, recuit simulé, recherche tabou, algo-
rithmes génétiques), de certaines moins connues (recherche à voisinages variables, GRASP,
iterated local search, guided local search, colonies de fourmis) et de techniques avanc ées (al-
gorithmes mémétiques, scatter search, GA|PM).
Pour toutes ces méthodes, nous analysons les facteurs d’intensification et de diversi-
fication présents, les particularités de chacune d’elle et notre retour d’expérience sur les
applications que nous avons traités. De cette analyse, nous pouvons proposer ce que sont,
selon nous, les caractéristiques indispensables à une bonne métaheuristique.
M ETAHEURISTICS :
STRATEGIES FOR THE OPTIMISATION OF THE PRODUCTION OF GOODS AND SERVICES
Solving optimisation problems is a key point in the constant improving of productivity in

the industries. When traditional methods fail, it is natural to look towards some approxi-
mate resolution methods.
Metaheuristics play, today, an important role in solving optmisation problems. Those
techniques became, in few years, necessary and powerful tools. In this synthesis, we present
an overview of the classical metaheuristics (descent methods, simulated annealing, tabu
search, genetic algorithms), some less known (variable neighbourhood search, GRASP, ite-
rated local search, guided local search, ant colonies) and some advanced techniques (me-
metic algorithms, scatter search, GA|PM).
For all of these methods, we analyse intensification and diversification factors, specific
features of each of them and our experiment on various applications we dealt with. From
this analysis, we suggest what are, from our point of view, the necessary caracteristics of a
good metaheuristic.
Marc Sevaux, Mai 2000

Marc.Sevaux@univ-valenciennes.fr

Met A Heuristics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Met A Heuristics

Uploaded by

Copyright:

Available Formats

Marc Sevaux

H ABILITATION À D IRIGER DES R ECHERCHES

Préparée au Laboratoire d’Automatique, de Mécanique d’informatique

Soutenue le 1er Juillet 2004 devant le jury composé de :

Président Prof. Christian Prins

3 Supervision de travaux d’étudiants 2e et 3e cycle 20

4 Administration et animation de la recherche 24

5 Visibilité, rayonnement et autres activités 27

6 Contrats, projets et financements 33

8 Liste des publications 48

2 Méthodes de recherche locale 59

3 Métaheuristiques à base de population 71

III Sélection de publications 109

1 European Journal of Operational Research (2003) 111

2 4OR (2004 à paraı̂tre) 112

3 Journal of Heuristics (2004 soumis) 112

4 Computers and Operations Research (2004 à paraı̂tre) 113

5 Naval Research Logistics (2003) 113

La rédaction de ce manuscrit est l’aboutissement de plusieurs années de tra-

1.2 Fonction actuelle

1.3 Prime d’encadrement doctorale et de recherche

1.4 Parcours et formation

Etude de deux problèmes d’optimisation en planification et ordonnance-

1994 – 1995 Université Pierre et Marie Curie, Paris

Année 2003-2004 Niveau CM TD TP EqTD

Année 2002-2003 Niveau CM TD TP EqTD

Année 2001-2002 Niveau CM TD TP EqTD

Année 2000-2001 Niveau CM TD TP EqTD

Année 1999-2000 Niveau CM TD TP EqTD

Descriptif des interventions à l’UVHC

(arbres programmatiques, langage structuré de 4e génération, concepts

2.2 Enseignements avant l’intégration à l’UVHC

2.3 Encadrements pédagogiques

Encadrements avant l’intégration à l’UVHC

2.4 Administration de l’enseignement

Responsable des relations internationales – depuis 2003

Responsable de l’organisation des projets – 2001-2003

Organisation des visites en entreprise – 1999-2001

Recrutement, admission, jury

3 Supervision de travaux d’étudiants 2e et 3e cycle

Frédéric BEUGNIES a débuté une thèse (bourse ministère) en septembre 2003

Co-direction de Thèse (2003-2006)

Co-direction de Thèse (2000-2004)

M.-J. Huguet, A. Thomas (Examinateurs).

3.2 Mémoires de DEA

DEA AISIH (2003-2004)

DEA AISIH (2002-2003)

DEA AISIH (2000-2001)

Christophe TILLEUL, étudiant en DEA a réalisé un stage au sein de notre

3.3 Projets de DESS

Chris BAUCHOT et Sébastien LEGENDRE ont réalisé au sein de notre équipe

DESS CCI (2002-2003)

Shiva ROUHOLAMINI et Rachid HARMAOUI, étudiants en DESS “Connais-

DESS ICHM (1999-2000)

3.4 Projets IUP GEII

IUP GEII (2001-2002)

IUP GEII (2001-2002)

4 Administration et animation de la recherche

Création et pilotage du groupe EU/ME (EURO Working Group – European chap-

Création et animation (de 2000 à 2002) des séminaires du LAMIH permettant à

Chargé de la collecte des informations (publications et activités de recherche et

4.2 Organisation de manifestations à Valenciennes

4.3 Organisation de manifestations en dehors de Valenciennes